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/SmallString.h" 17 #include "llvm/ADT/StringRef.h" 18 #include "llvm/ADT/TinyPtrVector.h" 19 #include "llvm/ADT/Twine.h" 20 #include "llvm/BinaryFormat/COFF.h" 21 #include "llvm/BinaryFormat/Dwarf.h" 22 #include "llvm/CodeGen/AsmPrinter.h" 23 #include "llvm/CodeGen/LexicalScopes.h" 24 #include "llvm/CodeGen/MachineFrameInfo.h" 25 #include "llvm/CodeGen/MachineFunction.h" 26 #include "llvm/CodeGen/MachineInstr.h" 27 #include "llvm/CodeGen/MachineModuleInfo.h" 28 #include "llvm/CodeGen/TargetFrameLowering.h" 29 #include "llvm/CodeGen/TargetLowering.h" 30 #include "llvm/CodeGen/TargetRegisterInfo.h" 31 #include "llvm/CodeGen/TargetSubtargetInfo.h" 32 #include "llvm/Config/llvm-config.h" 33 #include "llvm/DebugInfo/CodeView/CVTypeVisitor.h" 34 #include "llvm/DebugInfo/CodeView/CodeViewRecordIO.h" 35 #include "llvm/DebugInfo/CodeView/ContinuationRecordBuilder.h" 36 #include "llvm/DebugInfo/CodeView/DebugInlineeLinesSubsection.h" 37 #include "llvm/DebugInfo/CodeView/EnumTables.h" 38 #include "llvm/DebugInfo/CodeView/Line.h" 39 #include "llvm/DebugInfo/CodeView/SymbolRecord.h" 40 #include "llvm/DebugInfo/CodeView/TypeRecord.h" 41 #include "llvm/DebugInfo/CodeView/TypeTableCollection.h" 42 #include "llvm/DebugInfo/CodeView/TypeVisitorCallbackPipeline.h" 43 #include "llvm/IR/Constants.h" 44 #include "llvm/IR/DataLayout.h" 45 #include "llvm/IR/DebugInfoMetadata.h" 46 #include "llvm/IR/Function.h" 47 #include "llvm/IR/GlobalValue.h" 48 #include "llvm/IR/GlobalVariable.h" 49 #include "llvm/IR/Metadata.h" 50 #include "llvm/IR/Module.h" 51 #include "llvm/MC/MCAsmInfo.h" 52 #include "llvm/MC/MCContext.h" 53 #include "llvm/MC/MCSectionCOFF.h" 54 #include "llvm/MC/MCStreamer.h" 55 #include "llvm/MC/MCSymbol.h" 56 #include "llvm/Support/BinaryStreamWriter.h" 57 #include "llvm/Support/Casting.h" 58 #include "llvm/Support/Endian.h" 59 #include "llvm/Support/Error.h" 60 #include "llvm/Support/ErrorHandling.h" 61 #include "llvm/Support/FormatVariadic.h" 62 #include "llvm/Support/Path.h" 63 #include "llvm/Support/Program.h" 64 #include "llvm/Support/SMLoc.h" 65 #include "llvm/Support/ScopedPrinter.h" 66 #include "llvm/Target/TargetLoweringObjectFile.h" 67 #include "llvm/Target/TargetMachine.h" 68 #include "llvm/TargetParser/Triple.h" 69 #include <algorithm> 70 #include <cassert> 71 #include <cctype> 72 #include <cstddef> 73 #include <iterator> 74 #include <limits> 75 76 using namespace llvm; 77 using namespace llvm::codeview; 78 79 namespace { 80 class CVMCAdapter : public CodeViewRecordStreamer { 81 public: 82 CVMCAdapter(MCStreamer &OS, TypeCollection &TypeTable) 83 : OS(&OS), TypeTable(TypeTable) {} 84 85 void emitBytes(StringRef Data) override { OS->emitBytes(Data); } 86 87 void emitIntValue(uint64_t Value, unsigned Size) override { 88 OS->emitIntValueInHex(Value, Size); 89 } 90 91 void emitBinaryData(StringRef Data) override { OS->emitBinaryData(Data); } 92 93 void AddComment(const Twine &T) override { OS->AddComment(T); } 94 95 void AddRawComment(const Twine &T) override { OS->emitRawComment(T); } 96 97 bool isVerboseAsm() override { return OS->isVerboseAsm(); } 98 99 std::string getTypeName(TypeIndex TI) override { 100 std::string TypeName; 101 if (!TI.isNoneType()) { 102 if (TI.isSimple()) 103 TypeName = std::string(TypeIndex::simpleTypeName(TI)); 104 else 105 TypeName = std::string(TypeTable.getTypeName(TI)); 106 } 107 return TypeName; 108 } 109 110 private: 111 MCStreamer *OS = nullptr; 112 TypeCollection &TypeTable; 113 }; 114 } // namespace 115 116 static CPUType mapArchToCVCPUType(Triple::ArchType Type) { 117 switch (Type) { 118 case Triple::ArchType::x86: 119 return CPUType::Pentium3; 120 case Triple::ArchType::x86_64: 121 return CPUType::X64; 122 case Triple::ArchType::thumb: 123 // LLVM currently doesn't support Windows CE and so thumb 124 // here is indiscriminately mapped to ARMNT specifically. 125 return CPUType::ARMNT; 126 case Triple::ArchType::aarch64: 127 return CPUType::ARM64; 128 default: 129 report_fatal_error("target architecture doesn't map to a CodeView CPUType"); 130 } 131 } 132 133 CodeViewDebug::CodeViewDebug(AsmPrinter *AP) 134 : DebugHandlerBase(AP), OS(*Asm->OutStreamer), TypeTable(Allocator) {} 135 136 StringRef CodeViewDebug::getFullFilepath(const DIFile *File) { 137 std::string &Filepath = FileToFilepathMap[File]; 138 if (!Filepath.empty()) 139 return Filepath; 140 141 StringRef Dir = File->getDirectory(), Filename = File->getFilename(); 142 143 // If this is a Unix-style path, just use it as is. Don't try to canonicalize 144 // it textually because one of the path components could be a symlink. 145 if (Dir.startswith("/") || Filename.startswith("/")) { 146 if (llvm::sys::path::is_absolute(Filename, llvm::sys::path::Style::posix)) 147 return Filename; 148 Filepath = std::string(Dir); 149 if (Dir.back() != '/') 150 Filepath += '/'; 151 Filepath += Filename; 152 return Filepath; 153 } 154 155 // Clang emits directory and relative filename info into the IR, but CodeView 156 // operates on full paths. We could change Clang to emit full paths too, but 157 // that would increase the IR size and probably not needed for other users. 158 // For now, just concatenate and canonicalize the path here. 159 if (Filename.find(':') == 1) 160 Filepath = std::string(Filename); 161 else 162 Filepath = (Dir + "\\" + Filename).str(); 163 164 // Canonicalize the path. We have to do it textually because we may no longer 165 // have access the file in the filesystem. 166 // First, replace all slashes with backslashes. 167 std::replace(Filepath.begin(), Filepath.end(), '/', '\\'); 168 169 // Remove all "\.\" with "\". 170 size_t Cursor = 0; 171 while ((Cursor = Filepath.find("\\.\\", Cursor)) != std::string::npos) 172 Filepath.erase(Cursor, 2); 173 174 // Replace all "\XXX\..\" with "\". Don't try too hard though as the original 175 // path should be well-formatted, e.g. start with a drive letter, etc. 176 Cursor = 0; 177 while ((Cursor = Filepath.find("\\..\\", Cursor)) != std::string::npos) { 178 // Something's wrong if the path starts with "\..\", abort. 179 if (Cursor == 0) 180 break; 181 182 size_t PrevSlash = Filepath.rfind('\\', Cursor - 1); 183 if (PrevSlash == std::string::npos) 184 // Something's wrong, abort. 185 break; 186 187 Filepath.erase(PrevSlash, Cursor + 3 - PrevSlash); 188 // The next ".." might be following the one we've just erased. 189 Cursor = PrevSlash; 190 } 191 192 // Remove all duplicate backslashes. 193 Cursor = 0; 194 while ((Cursor = Filepath.find("\\\\", Cursor)) != std::string::npos) 195 Filepath.erase(Cursor, 1); 196 197 return Filepath; 198 } 199 200 unsigned CodeViewDebug::maybeRecordFile(const DIFile *F) { 201 StringRef FullPath = getFullFilepath(F); 202 unsigned NextId = FileIdMap.size() + 1; 203 auto Insertion = FileIdMap.insert(std::make_pair(FullPath, NextId)); 204 if (Insertion.second) { 205 // We have to compute the full filepath and emit a .cv_file directive. 206 ArrayRef<uint8_t> ChecksumAsBytes; 207 FileChecksumKind CSKind = FileChecksumKind::None; 208 if (F->getChecksum()) { 209 std::string Checksum = fromHex(F->getChecksum()->Value); 210 void *CKMem = OS.getContext().allocate(Checksum.size(), 1); 211 memcpy(CKMem, Checksum.data(), Checksum.size()); 212 ChecksumAsBytes = ArrayRef<uint8_t>( 213 reinterpret_cast<const uint8_t *>(CKMem), Checksum.size()); 214 switch (F->getChecksum()->Kind) { 215 case DIFile::CSK_MD5: 216 CSKind = FileChecksumKind::MD5; 217 break; 218 case DIFile::CSK_SHA1: 219 CSKind = FileChecksumKind::SHA1; 220 break; 221 case DIFile::CSK_SHA256: 222 CSKind = FileChecksumKind::SHA256; 223 break; 224 } 225 } 226 bool Success = OS.emitCVFileDirective(NextId, FullPath, ChecksumAsBytes, 227 static_cast<unsigned>(CSKind)); 228 (void)Success; 229 assert(Success && ".cv_file directive failed"); 230 } 231 return Insertion.first->second; 232 } 233 234 CodeViewDebug::InlineSite & 235 CodeViewDebug::getInlineSite(const DILocation *InlinedAt, 236 const DISubprogram *Inlinee) { 237 auto SiteInsertion = CurFn->InlineSites.insert({InlinedAt, InlineSite()}); 238 InlineSite *Site = &SiteInsertion.first->second; 239 if (SiteInsertion.second) { 240 unsigned ParentFuncId = CurFn->FuncId; 241 if (const DILocation *OuterIA = InlinedAt->getInlinedAt()) 242 ParentFuncId = 243 getInlineSite(OuterIA, InlinedAt->getScope()->getSubprogram()) 244 .SiteFuncId; 245 246 Site->SiteFuncId = NextFuncId++; 247 OS.emitCVInlineSiteIdDirective( 248 Site->SiteFuncId, ParentFuncId, maybeRecordFile(InlinedAt->getFile()), 249 InlinedAt->getLine(), InlinedAt->getColumn(), SMLoc()); 250 Site->Inlinee = Inlinee; 251 InlinedSubprograms.insert(Inlinee); 252 getFuncIdForSubprogram(Inlinee); 253 } 254 return *Site; 255 } 256 257 static StringRef getPrettyScopeName(const DIScope *Scope) { 258 StringRef ScopeName = Scope->getName(); 259 if (!ScopeName.empty()) 260 return ScopeName; 261 262 switch (Scope->getTag()) { 263 case dwarf::DW_TAG_enumeration_type: 264 case dwarf::DW_TAG_class_type: 265 case dwarf::DW_TAG_structure_type: 266 case dwarf::DW_TAG_union_type: 267 return "<unnamed-tag>"; 268 case dwarf::DW_TAG_namespace: 269 return "`anonymous namespace'"; 270 default: 271 return StringRef(); 272 } 273 } 274 275 const DISubprogram *CodeViewDebug::collectParentScopeNames( 276 const DIScope *Scope, SmallVectorImpl<StringRef> &QualifiedNameComponents) { 277 const DISubprogram *ClosestSubprogram = nullptr; 278 while (Scope != nullptr) { 279 if (ClosestSubprogram == nullptr) 280 ClosestSubprogram = dyn_cast<DISubprogram>(Scope); 281 282 // If a type appears in a scope chain, make sure it gets emitted. The 283 // frontend will be responsible for deciding if this should be a forward 284 // declaration or a complete type. 285 if (const auto *Ty = dyn_cast<DICompositeType>(Scope)) 286 DeferredCompleteTypes.push_back(Ty); 287 288 StringRef ScopeName = getPrettyScopeName(Scope); 289 if (!ScopeName.empty()) 290 QualifiedNameComponents.push_back(ScopeName); 291 Scope = Scope->getScope(); 292 } 293 return ClosestSubprogram; 294 } 295 296 static std::string formatNestedName(ArrayRef<StringRef> QualifiedNameComponents, 297 StringRef TypeName) { 298 std::string FullyQualifiedName; 299 for (StringRef QualifiedNameComponent : 300 llvm::reverse(QualifiedNameComponents)) { 301 FullyQualifiedName.append(std::string(QualifiedNameComponent)); 302 FullyQualifiedName.append("::"); 303 } 304 FullyQualifiedName.append(std::string(TypeName)); 305 return FullyQualifiedName; 306 } 307 308 struct CodeViewDebug::TypeLoweringScope { 309 TypeLoweringScope(CodeViewDebug &CVD) : CVD(CVD) { ++CVD.TypeEmissionLevel; } 310 ~TypeLoweringScope() { 311 // Don't decrement TypeEmissionLevel until after emitting deferred types, so 312 // inner TypeLoweringScopes don't attempt to emit deferred types. 313 if (CVD.TypeEmissionLevel == 1) 314 CVD.emitDeferredCompleteTypes(); 315 --CVD.TypeEmissionLevel; 316 } 317 CodeViewDebug &CVD; 318 }; 319 320 std::string CodeViewDebug::getFullyQualifiedName(const DIScope *Scope, 321 StringRef Name) { 322 // Ensure types in the scope chain are emitted as soon as possible. 323 // This can create otherwise a situation where S_UDTs are emitted while 324 // looping in emitDebugInfoForUDTs. 325 TypeLoweringScope S(*this); 326 SmallVector<StringRef, 5> QualifiedNameComponents; 327 collectParentScopeNames(Scope, QualifiedNameComponents); 328 return formatNestedName(QualifiedNameComponents, Name); 329 } 330 331 std::string CodeViewDebug::getFullyQualifiedName(const DIScope *Ty) { 332 const DIScope *Scope = Ty->getScope(); 333 return getFullyQualifiedName(Scope, getPrettyScopeName(Ty)); 334 } 335 336 TypeIndex CodeViewDebug::getScopeIndex(const DIScope *Scope) { 337 // No scope means global scope and that uses the zero index. 338 // 339 // We also use zero index when the scope is a DISubprogram 340 // to suppress the emission of LF_STRING_ID for the function, 341 // which can trigger a link-time error with the linker in 342 // VS2019 version 16.11.2 or newer. 343 // Note, however, skipping the debug info emission for the DISubprogram 344 // is a temporary fix. The root issue here is that we need to figure out 345 // the proper way to encode a function nested in another function 346 // (as introduced by the Fortran 'contains' keyword) in CodeView. 347 if (!Scope || isa<DIFile>(Scope) || isa<DISubprogram>(Scope)) 348 return TypeIndex(); 349 350 assert(!isa<DIType>(Scope) && "shouldn't make a namespace scope for a type"); 351 352 // Check if we've already translated this scope. 353 auto I = TypeIndices.find({Scope, nullptr}); 354 if (I != TypeIndices.end()) 355 return I->second; 356 357 // Build the fully qualified name of the scope. 358 std::string ScopeName = getFullyQualifiedName(Scope); 359 StringIdRecord SID(TypeIndex(), ScopeName); 360 auto TI = TypeTable.writeLeafType(SID); 361 return recordTypeIndexForDINode(Scope, TI); 362 } 363 364 static StringRef removeTemplateArgs(StringRef Name) { 365 // Remove template args from the display name. Assume that the template args 366 // are the last thing in the name. 367 if (Name.empty() || Name.back() != '>') 368 return Name; 369 370 int OpenBrackets = 0; 371 for (int i = Name.size() - 1; i >= 0; --i) { 372 if (Name[i] == '>') 373 ++OpenBrackets; 374 else if (Name[i] == '<') { 375 --OpenBrackets; 376 if (OpenBrackets == 0) 377 return Name.substr(0, i); 378 } 379 } 380 return Name; 381 } 382 383 TypeIndex CodeViewDebug::getFuncIdForSubprogram(const DISubprogram *SP) { 384 assert(SP); 385 386 // Check if we've already translated this subprogram. 387 auto I = TypeIndices.find({SP, nullptr}); 388 if (I != TypeIndices.end()) 389 return I->second; 390 391 // The display name includes function template arguments. Drop them to match 392 // MSVC. We need to have the template arguments in the DISubprogram name 393 // because they are used in other symbol records, such as S_GPROC32_IDs. 394 StringRef DisplayName = removeTemplateArgs(SP->getName()); 395 396 const DIScope *Scope = SP->getScope(); 397 TypeIndex TI; 398 if (const auto *Class = dyn_cast_or_null<DICompositeType>(Scope)) { 399 // If the scope is a DICompositeType, then this must be a method. Member 400 // function types take some special handling, and require access to the 401 // subprogram. 402 TypeIndex ClassType = getTypeIndex(Class); 403 MemberFuncIdRecord MFuncId(ClassType, getMemberFunctionType(SP, Class), 404 DisplayName); 405 TI = TypeTable.writeLeafType(MFuncId); 406 } else { 407 // Otherwise, this must be a free function. 408 TypeIndex ParentScope = getScopeIndex(Scope); 409 FuncIdRecord FuncId(ParentScope, getTypeIndex(SP->getType()), DisplayName); 410 TI = TypeTable.writeLeafType(FuncId); 411 } 412 413 return recordTypeIndexForDINode(SP, TI); 414 } 415 416 static bool isNonTrivial(const DICompositeType *DCTy) { 417 return ((DCTy->getFlags() & DINode::FlagNonTrivial) == DINode::FlagNonTrivial); 418 } 419 420 static FunctionOptions 421 getFunctionOptions(const DISubroutineType *Ty, 422 const DICompositeType *ClassTy = nullptr, 423 StringRef SPName = StringRef("")) { 424 FunctionOptions FO = FunctionOptions::None; 425 const DIType *ReturnTy = nullptr; 426 if (auto TypeArray = Ty->getTypeArray()) { 427 if (TypeArray.size()) 428 ReturnTy = TypeArray[0]; 429 } 430 431 // Add CxxReturnUdt option to functions that return nontrivial record types 432 // or methods that return record types. 433 if (auto *ReturnDCTy = dyn_cast_or_null<DICompositeType>(ReturnTy)) 434 if (isNonTrivial(ReturnDCTy) || ClassTy) 435 FO |= FunctionOptions::CxxReturnUdt; 436 437 // DISubroutineType is unnamed. Use DISubprogram's i.e. SPName in comparison. 438 if (ClassTy && isNonTrivial(ClassTy) && SPName == ClassTy->getName()) { 439 FO |= FunctionOptions::Constructor; 440 441 // TODO: put the FunctionOptions::ConstructorWithVirtualBases flag. 442 443 } 444 return FO; 445 } 446 447 TypeIndex CodeViewDebug::getMemberFunctionType(const DISubprogram *SP, 448 const DICompositeType *Class) { 449 // Always use the method declaration as the key for the function type. The 450 // method declaration contains the this adjustment. 451 if (SP->getDeclaration()) 452 SP = SP->getDeclaration(); 453 assert(!SP->getDeclaration() && "should use declaration as key"); 454 455 // Key the MemberFunctionRecord into the map as {SP, Class}. It won't collide 456 // with the MemberFuncIdRecord, which is keyed in as {SP, nullptr}. 457 auto I = TypeIndices.find({SP, Class}); 458 if (I != TypeIndices.end()) 459 return I->second; 460 461 // Make sure complete type info for the class is emitted *after* the member 462 // function type, as the complete class type is likely to reference this 463 // member function type. 464 TypeLoweringScope S(*this); 465 const bool IsStaticMethod = (SP->getFlags() & DINode::FlagStaticMember) != 0; 466 467 FunctionOptions FO = getFunctionOptions(SP->getType(), Class, SP->getName()); 468 TypeIndex TI = lowerTypeMemberFunction( 469 SP->getType(), Class, SP->getThisAdjustment(), IsStaticMethod, FO); 470 return recordTypeIndexForDINode(SP, TI, Class); 471 } 472 473 TypeIndex CodeViewDebug::recordTypeIndexForDINode(const DINode *Node, 474 TypeIndex TI, 475 const DIType *ClassTy) { 476 auto InsertResult = TypeIndices.insert({{Node, ClassTy}, TI}); 477 (void)InsertResult; 478 assert(InsertResult.second && "DINode was already assigned a type index"); 479 return TI; 480 } 481 482 unsigned CodeViewDebug::getPointerSizeInBytes() { 483 return MMI->getModule()->getDataLayout().getPointerSizeInBits() / 8; 484 } 485 486 void CodeViewDebug::recordLocalVariable(LocalVariable &&Var, 487 const LexicalScope *LS) { 488 if (const DILocation *InlinedAt = LS->getInlinedAt()) { 489 // This variable was inlined. Associate it with the InlineSite. 490 const DISubprogram *Inlinee = Var.DIVar->getScope()->getSubprogram(); 491 InlineSite &Site = getInlineSite(InlinedAt, Inlinee); 492 Site.InlinedLocals.emplace_back(std::move(Var)); 493 } else { 494 // This variable goes into the corresponding lexical scope. 495 ScopeVariables[LS].emplace_back(std::move(Var)); 496 } 497 } 498 499 static void addLocIfNotPresent(SmallVectorImpl<const DILocation *> &Locs, 500 const DILocation *Loc) { 501 if (!llvm::is_contained(Locs, Loc)) 502 Locs.push_back(Loc); 503 } 504 505 void CodeViewDebug::maybeRecordLocation(const DebugLoc &DL, 506 const MachineFunction *MF) { 507 // Skip this instruction if it has the same location as the previous one. 508 if (!DL || DL == PrevInstLoc) 509 return; 510 511 const DIScope *Scope = DL->getScope(); 512 if (!Scope) 513 return; 514 515 // Skip this line if it is longer than the maximum we can record. 516 LineInfo LI(DL.getLine(), DL.getLine(), /*IsStatement=*/true); 517 if (LI.getStartLine() != DL.getLine() || LI.isAlwaysStepInto() || 518 LI.isNeverStepInto()) 519 return; 520 521 ColumnInfo CI(DL.getCol(), /*EndColumn=*/0); 522 if (CI.getStartColumn() != DL.getCol()) 523 return; 524 525 if (!CurFn->HaveLineInfo) 526 CurFn->HaveLineInfo = true; 527 unsigned FileId = 0; 528 if (PrevInstLoc.get() && PrevInstLoc->getFile() == DL->getFile()) 529 FileId = CurFn->LastFileId; 530 else 531 FileId = CurFn->LastFileId = maybeRecordFile(DL->getFile()); 532 PrevInstLoc = DL; 533 534 unsigned FuncId = CurFn->FuncId; 535 if (const DILocation *SiteLoc = DL->getInlinedAt()) { 536 const DILocation *Loc = DL.get(); 537 538 // If this location was actually inlined from somewhere else, give it the ID 539 // of the inline call site. 540 FuncId = 541 getInlineSite(SiteLoc, Loc->getScope()->getSubprogram()).SiteFuncId; 542 543 // Ensure we have links in the tree of inline call sites. 544 bool FirstLoc = true; 545 while ((SiteLoc = Loc->getInlinedAt())) { 546 InlineSite &Site = 547 getInlineSite(SiteLoc, Loc->getScope()->getSubprogram()); 548 if (!FirstLoc) 549 addLocIfNotPresent(Site.ChildSites, Loc); 550 FirstLoc = false; 551 Loc = SiteLoc; 552 } 553 addLocIfNotPresent(CurFn->ChildSites, Loc); 554 } 555 556 OS.emitCVLocDirective(FuncId, FileId, DL.getLine(), DL.getCol(), 557 /*PrologueEnd=*/false, /*IsStmt=*/false, 558 DL->getFilename(), SMLoc()); 559 } 560 561 void CodeViewDebug::emitCodeViewMagicVersion() { 562 OS.emitValueToAlignment(Align(4)); 563 OS.AddComment("Debug section magic"); 564 OS.emitInt32(COFF::DEBUG_SECTION_MAGIC); 565 } 566 567 static SourceLanguage MapDWLangToCVLang(unsigned DWLang) { 568 switch (DWLang) { 569 case dwarf::DW_LANG_C: 570 case dwarf::DW_LANG_C89: 571 case dwarf::DW_LANG_C99: 572 case dwarf::DW_LANG_C11: 573 return SourceLanguage::C; 574 case dwarf::DW_LANG_C_plus_plus: 575 case dwarf::DW_LANG_C_plus_plus_03: 576 case dwarf::DW_LANG_C_plus_plus_11: 577 case dwarf::DW_LANG_C_plus_plus_14: 578 return SourceLanguage::Cpp; 579 case dwarf::DW_LANG_Fortran77: 580 case dwarf::DW_LANG_Fortran90: 581 case dwarf::DW_LANG_Fortran95: 582 case dwarf::DW_LANG_Fortran03: 583 case dwarf::DW_LANG_Fortran08: 584 return SourceLanguage::Fortran; 585 case dwarf::DW_LANG_Pascal83: 586 return SourceLanguage::Pascal; 587 case dwarf::DW_LANG_Cobol74: 588 case dwarf::DW_LANG_Cobol85: 589 return SourceLanguage::Cobol; 590 case dwarf::DW_LANG_Java: 591 return SourceLanguage::Java; 592 case dwarf::DW_LANG_D: 593 return SourceLanguage::D; 594 case dwarf::DW_LANG_Swift: 595 return SourceLanguage::Swift; 596 case dwarf::DW_LANG_Rust: 597 return SourceLanguage::Rust; 598 case dwarf::DW_LANG_ObjC: 599 return SourceLanguage::ObjC; 600 case dwarf::DW_LANG_ObjC_plus_plus: 601 return SourceLanguage::ObjCpp; 602 default: 603 // There's no CodeView representation for this language, and CV doesn't 604 // have an "unknown" option for the language field, so we'll use MASM, 605 // as it's very low level. 606 return SourceLanguage::Masm; 607 } 608 } 609 610 void CodeViewDebug::beginModule(Module *M) { 611 // If module doesn't have named metadata anchors or COFF debug section 612 // is not available, skip any debug info related stuff. 613 if (!MMI->hasDebugInfo() || 614 !Asm->getObjFileLowering().getCOFFDebugSymbolsSection()) { 615 Asm = nullptr; 616 return; 617 } 618 619 TheCPU = mapArchToCVCPUType(Triple(M->getTargetTriple()).getArch()); 620 621 // Get the current source language. 622 const MDNode *Node = *M->debug_compile_units_begin(); 623 const auto *CU = cast<DICompileUnit>(Node); 624 625 CurrentSourceLanguage = MapDWLangToCVLang(CU->getSourceLanguage()); 626 627 collectGlobalVariableInfo(); 628 629 // Check if we should emit type record hashes. 630 ConstantInt *GH = 631 mdconst::extract_or_null<ConstantInt>(M->getModuleFlag("CodeViewGHash")); 632 EmitDebugGlobalHashes = GH && !GH->isZero(); 633 } 634 635 void CodeViewDebug::endModule() { 636 if (!Asm || !MMI->hasDebugInfo()) 637 return; 638 639 // The COFF .debug$S section consists of several subsections, each starting 640 // with a 4-byte control code (e.g. 0xF1, 0xF2, etc) and then a 4-byte length 641 // of the payload followed by the payload itself. The subsections are 4-byte 642 // aligned. 643 644 // Use the generic .debug$S section, and make a subsection for all the inlined 645 // subprograms. 646 switchToDebugSectionForSymbol(nullptr); 647 648 MCSymbol *CompilerInfo = beginCVSubsection(DebugSubsectionKind::Symbols); 649 emitObjName(); 650 emitCompilerInformation(); 651 endCVSubsection(CompilerInfo); 652 653 emitInlineeLinesSubsection(); 654 655 // Emit per-function debug information. 656 for (auto &P : FnDebugInfo) 657 if (!P.first->isDeclarationForLinker()) 658 emitDebugInfoForFunction(P.first, *P.second); 659 660 // Get types used by globals without emitting anything. 661 // This is meant to collect all static const data members so they can be 662 // emitted as globals. 663 collectDebugInfoForGlobals(); 664 665 // Emit retained types. 666 emitDebugInfoForRetainedTypes(); 667 668 // Emit global variable debug information. 669 setCurrentSubprogram(nullptr); 670 emitDebugInfoForGlobals(); 671 672 // Switch back to the generic .debug$S section after potentially processing 673 // comdat symbol sections. 674 switchToDebugSectionForSymbol(nullptr); 675 676 // Emit UDT records for any types used by global variables. 677 if (!GlobalUDTs.empty()) { 678 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 679 emitDebugInfoForUDTs(GlobalUDTs); 680 endCVSubsection(SymbolsEnd); 681 } 682 683 // This subsection holds a file index to offset in string table table. 684 OS.AddComment("File index to string table offset subsection"); 685 OS.emitCVFileChecksumsDirective(); 686 687 // This subsection holds the string table. 688 OS.AddComment("String table"); 689 OS.emitCVStringTableDirective(); 690 691 // Emit S_BUILDINFO, which points to LF_BUILDINFO. Put this in its own symbol 692 // subsection in the generic .debug$S section at the end. There is no 693 // particular reason for this ordering other than to match MSVC. 694 emitBuildInfo(); 695 696 // Emit type information and hashes last, so that any types we translate while 697 // emitting function info are included. 698 emitTypeInformation(); 699 700 if (EmitDebugGlobalHashes) 701 emitTypeGlobalHashes(); 702 703 clear(); 704 } 705 706 static void 707 emitNullTerminatedSymbolName(MCStreamer &OS, StringRef S, 708 unsigned MaxFixedRecordLength = 0xF00) { 709 // The maximum CV record length is 0xFF00. Most of the strings we emit appear 710 // after a fixed length portion of the record. The fixed length portion should 711 // always be less than 0xF00 (3840) bytes, so truncate the string so that the 712 // overall record size is less than the maximum allowed. 713 SmallString<32> NullTerminatedString( 714 S.take_front(MaxRecordLength - MaxFixedRecordLength - 1)); 715 NullTerminatedString.push_back('\0'); 716 OS.emitBytes(NullTerminatedString); 717 } 718 719 void CodeViewDebug::emitTypeInformation() { 720 if (TypeTable.empty()) 721 return; 722 723 // Start the .debug$T or .debug$P section with 0x4. 724 OS.switchSection(Asm->getObjFileLowering().getCOFFDebugTypesSection()); 725 emitCodeViewMagicVersion(); 726 727 TypeTableCollection Table(TypeTable.records()); 728 TypeVisitorCallbackPipeline Pipeline; 729 730 // To emit type record using Codeview MCStreamer adapter 731 CVMCAdapter CVMCOS(OS, Table); 732 TypeRecordMapping typeMapping(CVMCOS); 733 Pipeline.addCallbackToPipeline(typeMapping); 734 735 std::optional<TypeIndex> B = Table.getFirst(); 736 while (B) { 737 // This will fail if the record data is invalid. 738 CVType Record = Table.getType(*B); 739 740 Error E = codeview::visitTypeRecord(Record, *B, Pipeline); 741 742 if (E) { 743 logAllUnhandledErrors(std::move(E), errs(), "error: "); 744 llvm_unreachable("produced malformed type record"); 745 } 746 747 B = Table.getNext(*B); 748 } 749 } 750 751 void CodeViewDebug::emitTypeGlobalHashes() { 752 if (TypeTable.empty()) 753 return; 754 755 // Start the .debug$H section with the version and hash algorithm, currently 756 // hardcoded to version 0, SHA1. 757 OS.switchSection(Asm->getObjFileLowering().getCOFFGlobalTypeHashesSection()); 758 759 OS.emitValueToAlignment(Align(4)); 760 OS.AddComment("Magic"); 761 OS.emitInt32(COFF::DEBUG_HASHES_SECTION_MAGIC); 762 OS.AddComment("Section Version"); 763 OS.emitInt16(0); 764 OS.AddComment("Hash Algorithm"); 765 OS.emitInt16(uint16_t(GlobalTypeHashAlg::BLAKE3)); 766 767 TypeIndex TI(TypeIndex::FirstNonSimpleIndex); 768 for (const auto &GHR : TypeTable.hashes()) { 769 if (OS.isVerboseAsm()) { 770 // Emit an EOL-comment describing which TypeIndex this hash corresponds 771 // to, as well as the stringified SHA1 hash. 772 SmallString<32> Comment; 773 raw_svector_ostream CommentOS(Comment); 774 CommentOS << formatv("{0:X+} [{1}]", TI.getIndex(), GHR); 775 OS.AddComment(Comment); 776 ++TI; 777 } 778 assert(GHR.Hash.size() == 8); 779 StringRef S(reinterpret_cast<const char *>(GHR.Hash.data()), 780 GHR.Hash.size()); 781 OS.emitBinaryData(S); 782 } 783 } 784 785 void CodeViewDebug::emitObjName() { 786 MCSymbol *CompilerEnd = beginSymbolRecord(SymbolKind::S_OBJNAME); 787 788 StringRef PathRef(Asm->TM.Options.ObjectFilenameForDebug); 789 llvm::SmallString<256> PathStore(PathRef); 790 791 if (PathRef.empty() || PathRef == "-") { 792 // Don't emit the filename if we're writing to stdout or to /dev/null. 793 PathRef = {}; 794 } else { 795 PathRef = PathStore; 796 } 797 798 OS.AddComment("Signature"); 799 OS.emitIntValue(0, 4); 800 801 OS.AddComment("Object name"); 802 emitNullTerminatedSymbolName(OS, PathRef); 803 804 endSymbolRecord(CompilerEnd); 805 } 806 807 namespace { 808 struct Version { 809 int Part[4]; 810 }; 811 } // end anonymous namespace 812 813 // Takes a StringRef like "clang 4.0.0.0 (other nonsense 123)" and parses out 814 // the version number. 815 static Version parseVersion(StringRef Name) { 816 Version V = {{0}}; 817 int N = 0; 818 for (const char C : Name) { 819 if (isdigit(C)) { 820 V.Part[N] *= 10; 821 V.Part[N] += C - '0'; 822 V.Part[N] = 823 std::min<int>(V.Part[N], std::numeric_limits<uint16_t>::max()); 824 } else if (C == '.') { 825 ++N; 826 if (N >= 4) 827 return V; 828 } else if (N > 0) 829 return V; 830 } 831 return V; 832 } 833 834 void CodeViewDebug::emitCompilerInformation() { 835 MCSymbol *CompilerEnd = beginSymbolRecord(SymbolKind::S_COMPILE3); 836 uint32_t Flags = 0; 837 838 // The low byte of the flags indicates the source language. 839 Flags = CurrentSourceLanguage; 840 // TODO: Figure out which other flags need to be set. 841 if (MMI->getModule()->getProfileSummary(/*IsCS*/ false) != nullptr) { 842 Flags |= static_cast<uint32_t>(CompileSym3Flags::PGO); 843 } 844 using ArchType = llvm::Triple::ArchType; 845 ArchType Arch = Triple(MMI->getModule()->getTargetTriple()).getArch(); 846 if (Asm->TM.Options.Hotpatch || Arch == ArchType::thumb || 847 Arch == ArchType::aarch64) { 848 Flags |= static_cast<uint32_t>(CompileSym3Flags::HotPatch); 849 } 850 851 OS.AddComment("Flags and language"); 852 OS.emitInt32(Flags); 853 854 OS.AddComment("CPUType"); 855 OS.emitInt16(static_cast<uint64_t>(TheCPU)); 856 857 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 858 const MDNode *Node = *CUs->operands().begin(); 859 const auto *CU = cast<DICompileUnit>(Node); 860 861 StringRef CompilerVersion = CU->getProducer(); 862 Version FrontVer = parseVersion(CompilerVersion); 863 OS.AddComment("Frontend version"); 864 for (int N : FrontVer.Part) { 865 OS.emitInt16(N); 866 } 867 868 // Some Microsoft tools, like Binscope, expect a backend version number of at 869 // least 8.something, so we'll coerce the LLVM version into a form that 870 // guarantees it'll be big enough without really lying about the version. 871 int Major = 1000 * LLVM_VERSION_MAJOR + 872 10 * LLVM_VERSION_MINOR + 873 LLVM_VERSION_PATCH; 874 // Clamp it for builds that use unusually large version numbers. 875 Major = std::min<int>(Major, std::numeric_limits<uint16_t>::max()); 876 Version BackVer = {{ Major, 0, 0, 0 }}; 877 OS.AddComment("Backend version"); 878 for (int N : BackVer.Part) 879 OS.emitInt16(N); 880 881 OS.AddComment("Null-terminated compiler version string"); 882 emitNullTerminatedSymbolName(OS, CompilerVersion); 883 884 endSymbolRecord(CompilerEnd); 885 } 886 887 static TypeIndex getStringIdTypeIdx(GlobalTypeTableBuilder &TypeTable, 888 StringRef S) { 889 StringIdRecord SIR(TypeIndex(0x0), S); 890 return TypeTable.writeLeafType(SIR); 891 } 892 893 static std::string flattenCommandLine(ArrayRef<std::string> Args, 894 StringRef MainFilename) { 895 std::string FlatCmdLine; 896 raw_string_ostream OS(FlatCmdLine); 897 bool PrintedOneArg = false; 898 if (!StringRef(Args[0]).contains("-cc1")) { 899 llvm::sys::printArg(OS, "-cc1", /*Quote=*/true); 900 PrintedOneArg = true; 901 } 902 for (unsigned i = 0; i < Args.size(); i++) { 903 StringRef Arg = Args[i]; 904 if (Arg.empty()) 905 continue; 906 if (Arg == "-main-file-name" || Arg == "-o") { 907 i++; // Skip this argument and next one. 908 continue; 909 } 910 if (Arg.startswith("-object-file-name") || Arg == MainFilename) 911 continue; 912 // Skip fmessage-length for reproduciability. 913 if (Arg.startswith("-fmessage-length")) 914 continue; 915 if (PrintedOneArg) 916 OS << " "; 917 llvm::sys::printArg(OS, Arg, /*Quote=*/true); 918 PrintedOneArg = true; 919 } 920 OS.flush(); 921 return FlatCmdLine; 922 } 923 924 void CodeViewDebug::emitBuildInfo() { 925 // First, make LF_BUILDINFO. It's a sequence of strings with various bits of 926 // build info. The known prefix is: 927 // - Absolute path of current directory 928 // - Compiler path 929 // - Main source file path, relative to CWD or absolute 930 // - Type server PDB file 931 // - Canonical compiler command line 932 // If frontend and backend compilation are separated (think llc or LTO), it's 933 // not clear if the compiler path should refer to the executable for the 934 // frontend or the backend. Leave it blank for now. 935 TypeIndex BuildInfoArgs[BuildInfoRecord::MaxArgs] = {}; 936 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 937 const MDNode *Node = *CUs->operands().begin(); // FIXME: Multiple CUs. 938 const auto *CU = cast<DICompileUnit>(Node); 939 const DIFile *MainSourceFile = CU->getFile(); 940 BuildInfoArgs[BuildInfoRecord::CurrentDirectory] = 941 getStringIdTypeIdx(TypeTable, MainSourceFile->getDirectory()); 942 BuildInfoArgs[BuildInfoRecord::SourceFile] = 943 getStringIdTypeIdx(TypeTable, MainSourceFile->getFilename()); 944 // FIXME: PDB is intentionally blank unless we implement /Zi type servers. 945 BuildInfoArgs[BuildInfoRecord::TypeServerPDB] = 946 getStringIdTypeIdx(TypeTable, ""); 947 if (Asm->TM.Options.MCOptions.Argv0 != nullptr) { 948 BuildInfoArgs[BuildInfoRecord::BuildTool] = 949 getStringIdTypeIdx(TypeTable, Asm->TM.Options.MCOptions.Argv0); 950 BuildInfoArgs[BuildInfoRecord::CommandLine] = getStringIdTypeIdx( 951 TypeTable, flattenCommandLine(Asm->TM.Options.MCOptions.CommandLineArgs, 952 MainSourceFile->getFilename())); 953 } 954 BuildInfoRecord BIR(BuildInfoArgs); 955 TypeIndex BuildInfoIndex = TypeTable.writeLeafType(BIR); 956 957 // Make a new .debug$S subsection for the S_BUILDINFO record, which points 958 // from the module symbols into the type stream. 959 MCSymbol *BISubsecEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 960 MCSymbol *BIEnd = beginSymbolRecord(SymbolKind::S_BUILDINFO); 961 OS.AddComment("LF_BUILDINFO index"); 962 OS.emitInt32(BuildInfoIndex.getIndex()); 963 endSymbolRecord(BIEnd); 964 endCVSubsection(BISubsecEnd); 965 } 966 967 void CodeViewDebug::emitInlineeLinesSubsection() { 968 if (InlinedSubprograms.empty()) 969 return; 970 971 OS.AddComment("Inlinee lines subsection"); 972 MCSymbol *InlineEnd = beginCVSubsection(DebugSubsectionKind::InlineeLines); 973 974 // We emit the checksum info for files. This is used by debuggers to 975 // determine if a pdb matches the source before loading it. Visual Studio, 976 // for instance, will display a warning that the breakpoints are not valid if 977 // the pdb does not match the source. 978 OS.AddComment("Inlinee lines signature"); 979 OS.emitInt32(unsigned(InlineeLinesSignature::Normal)); 980 981 for (const DISubprogram *SP : InlinedSubprograms) { 982 assert(TypeIndices.count({SP, nullptr})); 983 TypeIndex InlineeIdx = TypeIndices[{SP, nullptr}]; 984 985 OS.addBlankLine(); 986 unsigned FileId = maybeRecordFile(SP->getFile()); 987 OS.AddComment("Inlined function " + SP->getName() + " starts at " + 988 SP->getFilename() + Twine(':') + Twine(SP->getLine())); 989 OS.addBlankLine(); 990 OS.AddComment("Type index of inlined function"); 991 OS.emitInt32(InlineeIdx.getIndex()); 992 OS.AddComment("Offset into filechecksum table"); 993 OS.emitCVFileChecksumOffsetDirective(FileId); 994 OS.AddComment("Starting line number"); 995 OS.emitInt32(SP->getLine()); 996 } 997 998 endCVSubsection(InlineEnd); 999 } 1000 1001 void CodeViewDebug::emitInlinedCallSite(const FunctionInfo &FI, 1002 const DILocation *InlinedAt, 1003 const InlineSite &Site) { 1004 assert(TypeIndices.count({Site.Inlinee, nullptr})); 1005 TypeIndex InlineeIdx = TypeIndices[{Site.Inlinee, nullptr}]; 1006 1007 // SymbolRecord 1008 MCSymbol *InlineEnd = beginSymbolRecord(SymbolKind::S_INLINESITE); 1009 1010 OS.AddComment("PtrParent"); 1011 OS.emitInt32(0); 1012 OS.AddComment("PtrEnd"); 1013 OS.emitInt32(0); 1014 OS.AddComment("Inlinee type index"); 1015 OS.emitInt32(InlineeIdx.getIndex()); 1016 1017 unsigned FileId = maybeRecordFile(Site.Inlinee->getFile()); 1018 unsigned StartLineNum = Site.Inlinee->getLine(); 1019 1020 OS.emitCVInlineLinetableDirective(Site.SiteFuncId, FileId, StartLineNum, 1021 FI.Begin, FI.End); 1022 1023 endSymbolRecord(InlineEnd); 1024 1025 emitLocalVariableList(FI, Site.InlinedLocals); 1026 1027 // Recurse on child inlined call sites before closing the scope. 1028 for (const DILocation *ChildSite : Site.ChildSites) { 1029 auto I = FI.InlineSites.find(ChildSite); 1030 assert(I != FI.InlineSites.end() && 1031 "child site not in function inline site map"); 1032 emitInlinedCallSite(FI, ChildSite, I->second); 1033 } 1034 1035 // Close the scope. 1036 emitEndSymbolRecord(SymbolKind::S_INLINESITE_END); 1037 } 1038 1039 void CodeViewDebug::switchToDebugSectionForSymbol(const MCSymbol *GVSym) { 1040 // If we have a symbol, it may be in a section that is COMDAT. If so, find the 1041 // comdat key. A section may be comdat because of -ffunction-sections or 1042 // because it is comdat in the IR. 1043 MCSectionCOFF *GVSec = 1044 GVSym ? dyn_cast<MCSectionCOFF>(&GVSym->getSection()) : nullptr; 1045 const MCSymbol *KeySym = GVSec ? GVSec->getCOMDATSymbol() : nullptr; 1046 1047 MCSectionCOFF *DebugSec = cast<MCSectionCOFF>( 1048 Asm->getObjFileLowering().getCOFFDebugSymbolsSection()); 1049 DebugSec = OS.getContext().getAssociativeCOFFSection(DebugSec, KeySym); 1050 1051 OS.switchSection(DebugSec); 1052 1053 // Emit the magic version number if this is the first time we've switched to 1054 // this section. 1055 if (ComdatDebugSections.insert(DebugSec).second) 1056 emitCodeViewMagicVersion(); 1057 } 1058 1059 // Emit an S_THUNK32/S_END symbol pair for a thunk routine. 1060 // The only supported thunk ordinal is currently the standard type. 1061 void CodeViewDebug::emitDebugInfoForThunk(const Function *GV, 1062 FunctionInfo &FI, 1063 const MCSymbol *Fn) { 1064 std::string FuncName = 1065 std::string(GlobalValue::dropLLVMManglingEscape(GV->getName())); 1066 const ThunkOrdinal ordinal = ThunkOrdinal::Standard; // Only supported kind. 1067 1068 OS.AddComment("Symbol subsection for " + Twine(FuncName)); 1069 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 1070 1071 // Emit S_THUNK32 1072 MCSymbol *ThunkRecordEnd = beginSymbolRecord(SymbolKind::S_THUNK32); 1073 OS.AddComment("PtrParent"); 1074 OS.emitInt32(0); 1075 OS.AddComment("PtrEnd"); 1076 OS.emitInt32(0); 1077 OS.AddComment("PtrNext"); 1078 OS.emitInt32(0); 1079 OS.AddComment("Thunk section relative address"); 1080 OS.emitCOFFSecRel32(Fn, /*Offset=*/0); 1081 OS.AddComment("Thunk section index"); 1082 OS.emitCOFFSectionIndex(Fn); 1083 OS.AddComment("Code size"); 1084 OS.emitAbsoluteSymbolDiff(FI.End, Fn, 2); 1085 OS.AddComment("Ordinal"); 1086 OS.emitInt8(unsigned(ordinal)); 1087 OS.AddComment("Function name"); 1088 emitNullTerminatedSymbolName(OS, FuncName); 1089 // Additional fields specific to the thunk ordinal would go here. 1090 endSymbolRecord(ThunkRecordEnd); 1091 1092 // Local variables/inlined routines are purposely omitted here. The point of 1093 // marking this as a thunk is so Visual Studio will NOT stop in this routine. 1094 1095 // Emit S_PROC_ID_END 1096 emitEndSymbolRecord(SymbolKind::S_PROC_ID_END); 1097 1098 endCVSubsection(SymbolsEnd); 1099 } 1100 1101 void CodeViewDebug::emitDebugInfoForFunction(const Function *GV, 1102 FunctionInfo &FI) { 1103 // For each function there is a separate subsection which holds the PC to 1104 // file:line table. 1105 const MCSymbol *Fn = Asm->getSymbol(GV); 1106 assert(Fn); 1107 1108 // Switch to the to a comdat section, if appropriate. 1109 switchToDebugSectionForSymbol(Fn); 1110 1111 std::string FuncName; 1112 auto *SP = GV->getSubprogram(); 1113 assert(SP); 1114 setCurrentSubprogram(SP); 1115 1116 if (SP->isThunk()) { 1117 emitDebugInfoForThunk(GV, FI, Fn); 1118 return; 1119 } 1120 1121 // If we have a display name, build the fully qualified name by walking the 1122 // chain of scopes. 1123 if (!SP->getName().empty()) 1124 FuncName = getFullyQualifiedName(SP->getScope(), SP->getName()); 1125 1126 // If our DISubprogram name is empty, use the mangled name. 1127 if (FuncName.empty()) 1128 FuncName = std::string(GlobalValue::dropLLVMManglingEscape(GV->getName())); 1129 1130 // Emit FPO data, but only on 32-bit x86. No other platforms use it. 1131 if (Triple(MMI->getModule()->getTargetTriple()).getArch() == Triple::x86) 1132 OS.emitCVFPOData(Fn); 1133 1134 // Emit a symbol subsection, required by VS2012+ to find function boundaries. 1135 OS.AddComment("Symbol subsection for " + Twine(FuncName)); 1136 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 1137 { 1138 SymbolKind ProcKind = GV->hasLocalLinkage() ? SymbolKind::S_LPROC32_ID 1139 : SymbolKind::S_GPROC32_ID; 1140 MCSymbol *ProcRecordEnd = beginSymbolRecord(ProcKind); 1141 1142 // These fields are filled in by tools like CVPACK which run after the fact. 1143 OS.AddComment("PtrParent"); 1144 OS.emitInt32(0); 1145 OS.AddComment("PtrEnd"); 1146 OS.emitInt32(0); 1147 OS.AddComment("PtrNext"); 1148 OS.emitInt32(0); 1149 // This is the important bit that tells the debugger where the function 1150 // code is located and what's its size: 1151 OS.AddComment("Code size"); 1152 OS.emitAbsoluteSymbolDiff(FI.End, Fn, 4); 1153 OS.AddComment("Offset after prologue"); 1154 OS.emitInt32(0); 1155 OS.AddComment("Offset before epilogue"); 1156 OS.emitInt32(0); 1157 OS.AddComment("Function type index"); 1158 OS.emitInt32(getFuncIdForSubprogram(GV->getSubprogram()).getIndex()); 1159 OS.AddComment("Function section relative address"); 1160 OS.emitCOFFSecRel32(Fn, /*Offset=*/0); 1161 OS.AddComment("Function section index"); 1162 OS.emitCOFFSectionIndex(Fn); 1163 OS.AddComment("Flags"); 1164 ProcSymFlags ProcFlags = ProcSymFlags::HasOptimizedDebugInfo; 1165 if (FI.HasFramePointer) 1166 ProcFlags |= ProcSymFlags::HasFP; 1167 if (GV->hasFnAttribute(Attribute::NoReturn)) 1168 ProcFlags |= ProcSymFlags::IsNoReturn; 1169 if (GV->hasFnAttribute(Attribute::NoInline)) 1170 ProcFlags |= ProcSymFlags::IsNoInline; 1171 OS.emitInt8(static_cast<uint8_t>(ProcFlags)); 1172 // Emit the function display name as a null-terminated string. 1173 OS.AddComment("Function name"); 1174 // Truncate the name so we won't overflow the record length field. 1175 emitNullTerminatedSymbolName(OS, FuncName); 1176 endSymbolRecord(ProcRecordEnd); 1177 1178 MCSymbol *FrameProcEnd = beginSymbolRecord(SymbolKind::S_FRAMEPROC); 1179 // Subtract out the CSR size since MSVC excludes that and we include it. 1180 OS.AddComment("FrameSize"); 1181 OS.emitInt32(FI.FrameSize - FI.CSRSize); 1182 OS.AddComment("Padding"); 1183 OS.emitInt32(0); 1184 OS.AddComment("Offset of padding"); 1185 OS.emitInt32(0); 1186 OS.AddComment("Bytes of callee saved registers"); 1187 OS.emitInt32(FI.CSRSize); 1188 OS.AddComment("Exception handler offset"); 1189 OS.emitInt32(0); 1190 OS.AddComment("Exception handler section"); 1191 OS.emitInt16(0); 1192 OS.AddComment("Flags (defines frame register)"); 1193 OS.emitInt32(uint32_t(FI.FrameProcOpts)); 1194 endSymbolRecord(FrameProcEnd); 1195 1196 emitLocalVariableList(FI, FI.Locals); 1197 emitGlobalVariableList(FI.Globals); 1198 emitLexicalBlockList(FI.ChildBlocks, FI); 1199 1200 // Emit inlined call site information. Only emit functions inlined directly 1201 // into the parent function. We'll emit the other sites recursively as part 1202 // of their parent inline site. 1203 for (const DILocation *InlinedAt : FI.ChildSites) { 1204 auto I = FI.InlineSites.find(InlinedAt); 1205 assert(I != FI.InlineSites.end() && 1206 "child site not in function inline site map"); 1207 emitInlinedCallSite(FI, InlinedAt, I->second); 1208 } 1209 1210 for (auto Annot : FI.Annotations) { 1211 MCSymbol *Label = Annot.first; 1212 MDTuple *Strs = cast<MDTuple>(Annot.second); 1213 MCSymbol *AnnotEnd = beginSymbolRecord(SymbolKind::S_ANNOTATION); 1214 OS.emitCOFFSecRel32(Label, /*Offset=*/0); 1215 // FIXME: Make sure we don't overflow the max record size. 1216 OS.emitCOFFSectionIndex(Label); 1217 OS.emitInt16(Strs->getNumOperands()); 1218 for (Metadata *MD : Strs->operands()) { 1219 // MDStrings are null terminated, so we can do EmitBytes and get the 1220 // nice .asciz directive. 1221 StringRef Str = cast<MDString>(MD)->getString(); 1222 assert(Str.data()[Str.size()] == '\0' && "non-nullterminated MDString"); 1223 OS.emitBytes(StringRef(Str.data(), Str.size() + 1)); 1224 } 1225 endSymbolRecord(AnnotEnd); 1226 } 1227 1228 for (auto HeapAllocSite : FI.HeapAllocSites) { 1229 const MCSymbol *BeginLabel = std::get<0>(HeapAllocSite); 1230 const MCSymbol *EndLabel = std::get<1>(HeapAllocSite); 1231 const DIType *DITy = std::get<2>(HeapAllocSite); 1232 MCSymbol *HeapAllocEnd = beginSymbolRecord(SymbolKind::S_HEAPALLOCSITE); 1233 OS.AddComment("Call site offset"); 1234 OS.emitCOFFSecRel32(BeginLabel, /*Offset=*/0); 1235 OS.AddComment("Call site section index"); 1236 OS.emitCOFFSectionIndex(BeginLabel); 1237 OS.AddComment("Call instruction length"); 1238 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2); 1239 OS.AddComment("Type index"); 1240 OS.emitInt32(getCompleteTypeIndex(DITy).getIndex()); 1241 endSymbolRecord(HeapAllocEnd); 1242 } 1243 1244 if (SP != nullptr) 1245 emitDebugInfoForUDTs(LocalUDTs); 1246 1247 emitDebugInfoForJumpTables(FI); 1248 1249 // We're done with this function. 1250 emitEndSymbolRecord(SymbolKind::S_PROC_ID_END); 1251 } 1252 endCVSubsection(SymbolsEnd); 1253 1254 // We have an assembler directive that takes care of the whole line table. 1255 OS.emitCVLinetableDirective(FI.FuncId, Fn, FI.End); 1256 } 1257 1258 CodeViewDebug::LocalVarDef 1259 CodeViewDebug::createDefRangeMem(uint16_t CVRegister, int Offset) { 1260 LocalVarDef DR; 1261 DR.InMemory = -1; 1262 DR.DataOffset = Offset; 1263 assert(DR.DataOffset == Offset && "truncation"); 1264 DR.IsSubfield = 0; 1265 DR.StructOffset = 0; 1266 DR.CVRegister = CVRegister; 1267 return DR; 1268 } 1269 1270 void CodeViewDebug::collectVariableInfoFromMFTable( 1271 DenseSet<InlinedEntity> &Processed) { 1272 const MachineFunction &MF = *Asm->MF; 1273 const TargetSubtargetInfo &TSI = MF.getSubtarget(); 1274 const TargetFrameLowering *TFI = TSI.getFrameLowering(); 1275 const TargetRegisterInfo *TRI = TSI.getRegisterInfo(); 1276 1277 for (const MachineFunction::VariableDbgInfo &VI : 1278 MF.getInStackSlotVariableDbgInfo()) { 1279 if (!VI.Var) 1280 continue; 1281 assert(VI.Var->isValidLocationForIntrinsic(VI.Loc) && 1282 "Expected inlined-at fields to agree"); 1283 1284 Processed.insert(InlinedEntity(VI.Var, VI.Loc->getInlinedAt())); 1285 LexicalScope *Scope = LScopes.findLexicalScope(VI.Loc); 1286 1287 // If variable scope is not found then skip this variable. 1288 if (!Scope) 1289 continue; 1290 1291 // If the variable has an attached offset expression, extract it. 1292 // FIXME: Try to handle DW_OP_deref as well. 1293 int64_t ExprOffset = 0; 1294 bool Deref = false; 1295 if (VI.Expr) { 1296 // If there is one DW_OP_deref element, use offset of 0 and keep going. 1297 if (VI.Expr->getNumElements() == 1 && 1298 VI.Expr->getElement(0) == llvm::dwarf::DW_OP_deref) 1299 Deref = true; 1300 else if (!VI.Expr->extractIfOffset(ExprOffset)) 1301 continue; 1302 } 1303 1304 // Get the frame register used and the offset. 1305 Register FrameReg; 1306 StackOffset FrameOffset = 1307 TFI->getFrameIndexReference(*Asm->MF, VI.getStackSlot(), FrameReg); 1308 uint16_t CVReg = TRI->getCodeViewRegNum(FrameReg); 1309 1310 assert(!FrameOffset.getScalable() && 1311 "Frame offsets with a scalable component are not supported"); 1312 1313 // Calculate the label ranges. 1314 LocalVarDef DefRange = 1315 createDefRangeMem(CVReg, FrameOffset.getFixed() + ExprOffset); 1316 1317 LocalVariable Var; 1318 Var.DIVar = VI.Var; 1319 1320 for (const InsnRange &Range : Scope->getRanges()) { 1321 const MCSymbol *Begin = getLabelBeforeInsn(Range.first); 1322 const MCSymbol *End = getLabelAfterInsn(Range.second); 1323 End = End ? End : Asm->getFunctionEnd(); 1324 Var.DefRanges[DefRange].emplace_back(Begin, End); 1325 } 1326 1327 if (Deref) 1328 Var.UseReferenceType = true; 1329 1330 recordLocalVariable(std::move(Var), Scope); 1331 } 1332 } 1333 1334 static bool canUseReferenceType(const DbgVariableLocation &Loc) { 1335 return !Loc.LoadChain.empty() && Loc.LoadChain.back() == 0; 1336 } 1337 1338 static bool needsReferenceType(const DbgVariableLocation &Loc) { 1339 return Loc.LoadChain.size() == 2 && Loc.LoadChain.back() == 0; 1340 } 1341 1342 void CodeViewDebug::calculateRanges( 1343 LocalVariable &Var, const DbgValueHistoryMap::Entries &Entries) { 1344 const TargetRegisterInfo *TRI = Asm->MF->getSubtarget().getRegisterInfo(); 1345 1346 // Calculate the definition ranges. 1347 for (auto I = Entries.begin(), E = Entries.end(); I != E; ++I) { 1348 const auto &Entry = *I; 1349 if (!Entry.isDbgValue()) 1350 continue; 1351 const MachineInstr *DVInst = Entry.getInstr(); 1352 assert(DVInst->isDebugValue() && "Invalid History entry"); 1353 // FIXME: Find a way to represent constant variables, since they are 1354 // relatively common. 1355 std::optional<DbgVariableLocation> Location = 1356 DbgVariableLocation::extractFromMachineInstruction(*DVInst); 1357 if (!Location) 1358 { 1359 // When we don't have a location this is usually because LLVM has 1360 // transformed it into a constant and we only have an llvm.dbg.value. We 1361 // can't represent these well in CodeView since S_LOCAL only works on 1362 // registers and memory locations. Instead, we will pretend this to be a 1363 // constant value to at least have it show up in the debugger. 1364 auto Op = DVInst->getDebugOperand(0); 1365 if (Op.isImm()) 1366 Var.ConstantValue = APSInt(APInt(64, Op.getImm()), false); 1367 continue; 1368 } 1369 1370 // CodeView can only express variables in register and variables in memory 1371 // at a constant offset from a register. However, for variables passed 1372 // indirectly by pointer, it is common for that pointer to be spilled to a 1373 // stack location. For the special case of one offseted load followed by a 1374 // zero offset load (a pointer spilled to the stack), we change the type of 1375 // the local variable from a value type to a reference type. This tricks the 1376 // debugger into doing the load for us. 1377 if (Var.UseReferenceType) { 1378 // We're using a reference type. Drop the last zero offset load. 1379 if (canUseReferenceType(*Location)) 1380 Location->LoadChain.pop_back(); 1381 else 1382 continue; 1383 } else if (needsReferenceType(*Location)) { 1384 // This location can't be expressed without switching to a reference type. 1385 // Start over using that. 1386 Var.UseReferenceType = true; 1387 Var.DefRanges.clear(); 1388 calculateRanges(Var, Entries); 1389 return; 1390 } 1391 1392 // We can only handle a register or an offseted load of a register. 1393 if (Location->Register == 0 || Location->LoadChain.size() > 1) 1394 continue; 1395 1396 LocalVarDef DR; 1397 DR.CVRegister = TRI->getCodeViewRegNum(Location->Register); 1398 DR.InMemory = !Location->LoadChain.empty(); 1399 DR.DataOffset = 1400 !Location->LoadChain.empty() ? Location->LoadChain.back() : 0; 1401 if (Location->FragmentInfo) { 1402 DR.IsSubfield = true; 1403 DR.StructOffset = Location->FragmentInfo->OffsetInBits / 8; 1404 } else { 1405 DR.IsSubfield = false; 1406 DR.StructOffset = 0; 1407 } 1408 1409 // Compute the label range. 1410 const MCSymbol *Begin = getLabelBeforeInsn(Entry.getInstr()); 1411 const MCSymbol *End; 1412 if (Entry.getEndIndex() != DbgValueHistoryMap::NoEntry) { 1413 auto &EndingEntry = Entries[Entry.getEndIndex()]; 1414 End = EndingEntry.isDbgValue() 1415 ? getLabelBeforeInsn(EndingEntry.getInstr()) 1416 : getLabelAfterInsn(EndingEntry.getInstr()); 1417 } else 1418 End = Asm->getFunctionEnd(); 1419 1420 // If the last range end is our begin, just extend the last range. 1421 // Otherwise make a new range. 1422 SmallVectorImpl<std::pair<const MCSymbol *, const MCSymbol *>> &R = 1423 Var.DefRanges[DR]; 1424 if (!R.empty() && R.back().second == Begin) 1425 R.back().second = End; 1426 else 1427 R.emplace_back(Begin, End); 1428 1429 // FIXME: Do more range combining. 1430 } 1431 } 1432 1433 void CodeViewDebug::collectVariableInfo(const DISubprogram *SP) { 1434 DenseSet<InlinedEntity> Processed; 1435 // Grab the variable info that was squirreled away in the MMI side-table. 1436 collectVariableInfoFromMFTable(Processed); 1437 1438 for (const auto &I : DbgValues) { 1439 InlinedEntity IV = I.first; 1440 if (Processed.count(IV)) 1441 continue; 1442 const DILocalVariable *DIVar = cast<DILocalVariable>(IV.first); 1443 const DILocation *InlinedAt = IV.second; 1444 1445 // Instruction ranges, specifying where IV is accessible. 1446 const auto &Entries = I.second; 1447 1448 LexicalScope *Scope = nullptr; 1449 if (InlinedAt) 1450 Scope = LScopes.findInlinedScope(DIVar->getScope(), InlinedAt); 1451 else 1452 Scope = LScopes.findLexicalScope(DIVar->getScope()); 1453 // If variable scope is not found then skip this variable. 1454 if (!Scope) 1455 continue; 1456 1457 LocalVariable Var; 1458 Var.DIVar = DIVar; 1459 1460 calculateRanges(Var, Entries); 1461 recordLocalVariable(std::move(Var), Scope); 1462 } 1463 } 1464 1465 void CodeViewDebug::beginFunctionImpl(const MachineFunction *MF) { 1466 const TargetSubtargetInfo &TSI = MF->getSubtarget(); 1467 const TargetRegisterInfo *TRI = TSI.getRegisterInfo(); 1468 const MachineFrameInfo &MFI = MF->getFrameInfo(); 1469 const Function &GV = MF->getFunction(); 1470 auto Insertion = FnDebugInfo.insert({&GV, std::make_unique<FunctionInfo>()}); 1471 assert(Insertion.second && "function already has info"); 1472 CurFn = Insertion.first->second.get(); 1473 CurFn->FuncId = NextFuncId++; 1474 CurFn->Begin = Asm->getFunctionBegin(); 1475 1476 // The S_FRAMEPROC record reports the stack size, and how many bytes of 1477 // callee-saved registers were used. For targets that don't use a PUSH 1478 // instruction (AArch64), this will be zero. 1479 CurFn->CSRSize = MFI.getCVBytesOfCalleeSavedRegisters(); 1480 CurFn->FrameSize = MFI.getStackSize(); 1481 CurFn->OffsetAdjustment = MFI.getOffsetAdjustment(); 1482 CurFn->HasStackRealignment = TRI->hasStackRealignment(*MF); 1483 1484 // For this function S_FRAMEPROC record, figure out which codeview register 1485 // will be the frame pointer. 1486 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::None; // None. 1487 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::None; // None. 1488 if (CurFn->FrameSize > 0) { 1489 if (!TSI.getFrameLowering()->hasFP(*MF)) { 1490 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr; 1491 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::StackPtr; 1492 } else { 1493 CurFn->HasFramePointer = true; 1494 // If there is an FP, parameters are always relative to it. 1495 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::FramePtr; 1496 if (CurFn->HasStackRealignment) { 1497 // If the stack needs realignment, locals are relative to SP or VFRAME. 1498 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr; 1499 } else { 1500 // Otherwise, locals are relative to EBP, and we probably have VLAs or 1501 // other stack adjustments. 1502 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::FramePtr; 1503 } 1504 } 1505 } 1506 1507 // Compute other frame procedure options. 1508 FrameProcedureOptions FPO = FrameProcedureOptions::None; 1509 if (MFI.hasVarSizedObjects()) 1510 FPO |= FrameProcedureOptions::HasAlloca; 1511 if (MF->exposesReturnsTwice()) 1512 FPO |= FrameProcedureOptions::HasSetJmp; 1513 // FIXME: Set HasLongJmp if we ever track that info. 1514 if (MF->hasInlineAsm()) 1515 FPO |= FrameProcedureOptions::HasInlineAssembly; 1516 if (GV.hasPersonalityFn()) { 1517 if (isAsynchronousEHPersonality( 1518 classifyEHPersonality(GV.getPersonalityFn()))) 1519 FPO |= FrameProcedureOptions::HasStructuredExceptionHandling; 1520 else 1521 FPO |= FrameProcedureOptions::HasExceptionHandling; 1522 } 1523 if (GV.hasFnAttribute(Attribute::InlineHint)) 1524 FPO |= FrameProcedureOptions::MarkedInline; 1525 if (GV.hasFnAttribute(Attribute::Naked)) 1526 FPO |= FrameProcedureOptions::Naked; 1527 if (MFI.hasStackProtectorIndex()) { 1528 FPO |= FrameProcedureOptions::SecurityChecks; 1529 if (GV.hasFnAttribute(Attribute::StackProtectStrong) || 1530 GV.hasFnAttribute(Attribute::StackProtectReq)) { 1531 FPO |= FrameProcedureOptions::StrictSecurityChecks; 1532 } 1533 } else if (!GV.hasStackProtectorFnAttr()) { 1534 // __declspec(safebuffers) disables stack guards. 1535 FPO |= FrameProcedureOptions::SafeBuffers; 1536 } 1537 FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedLocalFramePtrReg) << 14U); 1538 FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedParamFramePtrReg) << 16U); 1539 if (Asm->TM.getOptLevel() != CodeGenOpt::None && 1540 !GV.hasOptSize() && !GV.hasOptNone()) 1541 FPO |= FrameProcedureOptions::OptimizedForSpeed; 1542 if (GV.hasProfileData()) { 1543 FPO |= FrameProcedureOptions::ValidProfileCounts; 1544 FPO |= FrameProcedureOptions::ProfileGuidedOptimization; 1545 } 1546 // FIXME: Set GuardCfg when it is implemented. 1547 CurFn->FrameProcOpts = FPO; 1548 1549 OS.emitCVFuncIdDirective(CurFn->FuncId); 1550 1551 // Find the end of the function prolog. First known non-DBG_VALUE and 1552 // non-frame setup location marks the beginning of the function body. 1553 // FIXME: is there a simpler a way to do this? Can we just search 1554 // for the first instruction of the function, not the last of the prolog? 1555 DebugLoc PrologEndLoc; 1556 bool EmptyPrologue = true; 1557 for (const auto &MBB : *MF) { 1558 for (const auto &MI : MBB) { 1559 if (!MI.isMetaInstruction() && !MI.getFlag(MachineInstr::FrameSetup) && 1560 MI.getDebugLoc()) { 1561 PrologEndLoc = MI.getDebugLoc(); 1562 break; 1563 } else if (!MI.isMetaInstruction()) { 1564 EmptyPrologue = false; 1565 } 1566 } 1567 } 1568 1569 // Record beginning of function if we have a non-empty prologue. 1570 if (PrologEndLoc && !EmptyPrologue) { 1571 DebugLoc FnStartDL = PrologEndLoc.getFnDebugLoc(); 1572 maybeRecordLocation(FnStartDL, MF); 1573 } 1574 1575 // Find heap alloc sites and emit labels around them. 1576 for (const auto &MBB : *MF) { 1577 for (const auto &MI : MBB) { 1578 if (MI.getHeapAllocMarker()) { 1579 requestLabelBeforeInsn(&MI); 1580 requestLabelAfterInsn(&MI); 1581 } 1582 } 1583 } 1584 1585 // Mark branches that may potentially be using jump tables with labels. 1586 bool isThumb = Triple(MMI->getModule()->getTargetTriple()).getArch() == 1587 llvm::Triple::ArchType::thumb; 1588 discoverJumpTableBranches(MF, isThumb); 1589 } 1590 1591 static bool shouldEmitUdt(const DIType *T) { 1592 if (!T) 1593 return false; 1594 1595 // MSVC does not emit UDTs for typedefs that are scoped to classes. 1596 if (T->getTag() == dwarf::DW_TAG_typedef) { 1597 if (DIScope *Scope = T->getScope()) { 1598 switch (Scope->getTag()) { 1599 case dwarf::DW_TAG_structure_type: 1600 case dwarf::DW_TAG_class_type: 1601 case dwarf::DW_TAG_union_type: 1602 return false; 1603 default: 1604 // do nothing. 1605 ; 1606 } 1607 } 1608 } 1609 1610 while (true) { 1611 if (!T || T->isForwardDecl()) 1612 return false; 1613 1614 const DIDerivedType *DT = dyn_cast<DIDerivedType>(T); 1615 if (!DT) 1616 return true; 1617 T = DT->getBaseType(); 1618 } 1619 return true; 1620 } 1621 1622 void CodeViewDebug::addToUDTs(const DIType *Ty) { 1623 // Don't record empty UDTs. 1624 if (Ty->getName().empty()) 1625 return; 1626 if (!shouldEmitUdt(Ty)) 1627 return; 1628 1629 SmallVector<StringRef, 5> ParentScopeNames; 1630 const DISubprogram *ClosestSubprogram = 1631 collectParentScopeNames(Ty->getScope(), ParentScopeNames); 1632 1633 std::string FullyQualifiedName = 1634 formatNestedName(ParentScopeNames, getPrettyScopeName(Ty)); 1635 1636 if (ClosestSubprogram == nullptr) { 1637 GlobalUDTs.emplace_back(std::move(FullyQualifiedName), Ty); 1638 } else if (ClosestSubprogram == CurrentSubprogram) { 1639 LocalUDTs.emplace_back(std::move(FullyQualifiedName), Ty); 1640 } 1641 1642 // TODO: What if the ClosestSubprogram is neither null or the current 1643 // subprogram? Currently, the UDT just gets dropped on the floor. 1644 // 1645 // The current behavior is not desirable. To get maximal fidelity, we would 1646 // need to perform all type translation before beginning emission of .debug$S 1647 // and then make LocalUDTs a member of FunctionInfo 1648 } 1649 1650 TypeIndex CodeViewDebug::lowerType(const DIType *Ty, const DIType *ClassTy) { 1651 // Generic dispatch for lowering an unknown type. 1652 switch (Ty->getTag()) { 1653 case dwarf::DW_TAG_array_type: 1654 return lowerTypeArray(cast<DICompositeType>(Ty)); 1655 case dwarf::DW_TAG_typedef: 1656 return lowerTypeAlias(cast<DIDerivedType>(Ty)); 1657 case dwarf::DW_TAG_base_type: 1658 return lowerTypeBasic(cast<DIBasicType>(Ty)); 1659 case dwarf::DW_TAG_pointer_type: 1660 if (cast<DIDerivedType>(Ty)->getName() == "__vtbl_ptr_type") 1661 return lowerTypeVFTableShape(cast<DIDerivedType>(Ty)); 1662 [[fallthrough]]; 1663 case dwarf::DW_TAG_reference_type: 1664 case dwarf::DW_TAG_rvalue_reference_type: 1665 return lowerTypePointer(cast<DIDerivedType>(Ty)); 1666 case dwarf::DW_TAG_ptr_to_member_type: 1667 return lowerTypeMemberPointer(cast<DIDerivedType>(Ty)); 1668 case dwarf::DW_TAG_restrict_type: 1669 case dwarf::DW_TAG_const_type: 1670 case dwarf::DW_TAG_volatile_type: 1671 // TODO: add support for DW_TAG_atomic_type here 1672 return lowerTypeModifier(cast<DIDerivedType>(Ty)); 1673 case dwarf::DW_TAG_subroutine_type: 1674 if (ClassTy) { 1675 // The member function type of a member function pointer has no 1676 // ThisAdjustment. 1677 return lowerTypeMemberFunction(cast<DISubroutineType>(Ty), ClassTy, 1678 /*ThisAdjustment=*/0, 1679 /*IsStaticMethod=*/false); 1680 } 1681 return lowerTypeFunction(cast<DISubroutineType>(Ty)); 1682 case dwarf::DW_TAG_enumeration_type: 1683 return lowerTypeEnum(cast<DICompositeType>(Ty)); 1684 case dwarf::DW_TAG_class_type: 1685 case dwarf::DW_TAG_structure_type: 1686 return lowerTypeClass(cast<DICompositeType>(Ty)); 1687 case dwarf::DW_TAG_union_type: 1688 return lowerTypeUnion(cast<DICompositeType>(Ty)); 1689 case dwarf::DW_TAG_string_type: 1690 return lowerTypeString(cast<DIStringType>(Ty)); 1691 case dwarf::DW_TAG_unspecified_type: 1692 if (Ty->getName() == "decltype(nullptr)") 1693 return TypeIndex::NullptrT(); 1694 return TypeIndex::None(); 1695 default: 1696 // Use the null type index. 1697 return TypeIndex(); 1698 } 1699 } 1700 1701 TypeIndex CodeViewDebug::lowerTypeAlias(const DIDerivedType *Ty) { 1702 TypeIndex UnderlyingTypeIndex = getTypeIndex(Ty->getBaseType()); 1703 StringRef TypeName = Ty->getName(); 1704 1705 addToUDTs(Ty); 1706 1707 if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::Int32Long) && 1708 TypeName == "HRESULT") 1709 return TypeIndex(SimpleTypeKind::HResult); 1710 if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::UInt16Short) && 1711 TypeName == "wchar_t") 1712 return TypeIndex(SimpleTypeKind::WideCharacter); 1713 1714 return UnderlyingTypeIndex; 1715 } 1716 1717 TypeIndex CodeViewDebug::lowerTypeArray(const DICompositeType *Ty) { 1718 const DIType *ElementType = Ty->getBaseType(); 1719 TypeIndex ElementTypeIndex = getTypeIndex(ElementType); 1720 // IndexType is size_t, which depends on the bitness of the target. 1721 TypeIndex IndexType = getPointerSizeInBytes() == 8 1722 ? TypeIndex(SimpleTypeKind::UInt64Quad) 1723 : TypeIndex(SimpleTypeKind::UInt32Long); 1724 1725 uint64_t ElementSize = getBaseTypeSize(ElementType) / 8; 1726 1727 // Add subranges to array type. 1728 DINodeArray Elements = Ty->getElements(); 1729 for (int i = Elements.size() - 1; i >= 0; --i) { 1730 const DINode *Element = Elements[i]; 1731 assert(Element->getTag() == dwarf::DW_TAG_subrange_type); 1732 1733 const DISubrange *Subrange = cast<DISubrange>(Element); 1734 int64_t Count = -1; 1735 1736 // If Subrange has a Count field, use it. 1737 // Otherwise, if it has an upperboud, use (upperbound - lowerbound + 1), 1738 // where lowerbound is from the LowerBound field of the Subrange, 1739 // or the language default lowerbound if that field is unspecified. 1740 if (auto *CI = dyn_cast_if_present<ConstantInt *>(Subrange->getCount())) 1741 Count = CI->getSExtValue(); 1742 else if (auto *UI = dyn_cast_if_present<ConstantInt *>( 1743 Subrange->getUpperBound())) { 1744 // Fortran uses 1 as the default lowerbound; other languages use 0. 1745 int64_t Lowerbound = (moduleIsInFortran()) ? 1 : 0; 1746 auto *LI = dyn_cast_if_present<ConstantInt *>(Subrange->getLowerBound()); 1747 Lowerbound = (LI) ? LI->getSExtValue() : Lowerbound; 1748 Count = UI->getSExtValue() - Lowerbound + 1; 1749 } 1750 1751 // Forward declarations of arrays without a size and VLAs use a count of -1. 1752 // Emit a count of zero in these cases to match what MSVC does for arrays 1753 // without a size. MSVC doesn't support VLAs, so it's not clear what we 1754 // should do for them even if we could distinguish them. 1755 if (Count == -1) 1756 Count = 0; 1757 1758 // Update the element size and element type index for subsequent subranges. 1759 ElementSize *= Count; 1760 1761 // If this is the outermost array, use the size from the array. It will be 1762 // more accurate if we had a VLA or an incomplete element type size. 1763 uint64_t ArraySize = 1764 (i == 0 && ElementSize == 0) ? Ty->getSizeInBits() / 8 : ElementSize; 1765 1766 StringRef Name = (i == 0) ? Ty->getName() : ""; 1767 ArrayRecord AR(ElementTypeIndex, IndexType, ArraySize, Name); 1768 ElementTypeIndex = TypeTable.writeLeafType(AR); 1769 } 1770 1771 return ElementTypeIndex; 1772 } 1773 1774 // This function lowers a Fortran character type (DIStringType). 1775 // Note that it handles only the character*n variant (using SizeInBits 1776 // field in DIString to describe the type size) at the moment. 1777 // Other variants (leveraging the StringLength and StringLengthExp 1778 // fields in DIStringType) remain TBD. 1779 TypeIndex CodeViewDebug::lowerTypeString(const DIStringType *Ty) { 1780 TypeIndex CharType = TypeIndex(SimpleTypeKind::NarrowCharacter); 1781 uint64_t ArraySize = Ty->getSizeInBits() >> 3; 1782 StringRef Name = Ty->getName(); 1783 // IndexType is size_t, which depends on the bitness of the target. 1784 TypeIndex IndexType = getPointerSizeInBytes() == 8 1785 ? TypeIndex(SimpleTypeKind::UInt64Quad) 1786 : TypeIndex(SimpleTypeKind::UInt32Long); 1787 1788 // Create a type of character array of ArraySize. 1789 ArrayRecord AR(CharType, IndexType, ArraySize, Name); 1790 1791 return TypeTable.writeLeafType(AR); 1792 } 1793 1794 TypeIndex CodeViewDebug::lowerTypeBasic(const DIBasicType *Ty) { 1795 TypeIndex Index; 1796 dwarf::TypeKind Kind; 1797 uint32_t ByteSize; 1798 1799 Kind = static_cast<dwarf::TypeKind>(Ty->getEncoding()); 1800 ByteSize = Ty->getSizeInBits() / 8; 1801 1802 SimpleTypeKind STK = SimpleTypeKind::None; 1803 switch (Kind) { 1804 case dwarf::DW_ATE_address: 1805 // FIXME: Translate 1806 break; 1807 case dwarf::DW_ATE_boolean: 1808 switch (ByteSize) { 1809 case 1: STK = SimpleTypeKind::Boolean8; break; 1810 case 2: STK = SimpleTypeKind::Boolean16; break; 1811 case 4: STK = SimpleTypeKind::Boolean32; break; 1812 case 8: STK = SimpleTypeKind::Boolean64; break; 1813 case 16: STK = SimpleTypeKind::Boolean128; break; 1814 } 1815 break; 1816 case dwarf::DW_ATE_complex_float: 1817 // The CodeView size for a complex represents the size of 1818 // an individual component. 1819 switch (ByteSize) { 1820 case 4: STK = SimpleTypeKind::Complex16; break; 1821 case 8: STK = SimpleTypeKind::Complex32; break; 1822 case 16: STK = SimpleTypeKind::Complex64; break; 1823 case 20: STK = SimpleTypeKind::Complex80; break; 1824 case 32: STK = SimpleTypeKind::Complex128; break; 1825 } 1826 break; 1827 case dwarf::DW_ATE_float: 1828 switch (ByteSize) { 1829 case 2: STK = SimpleTypeKind::Float16; break; 1830 case 4: STK = SimpleTypeKind::Float32; break; 1831 case 6: STK = SimpleTypeKind::Float48; break; 1832 case 8: STK = SimpleTypeKind::Float64; break; 1833 case 10: STK = SimpleTypeKind::Float80; break; 1834 case 16: STK = SimpleTypeKind::Float128; break; 1835 } 1836 break; 1837 case dwarf::DW_ATE_signed: 1838 switch (ByteSize) { 1839 case 1: STK = SimpleTypeKind::SignedCharacter; break; 1840 case 2: STK = SimpleTypeKind::Int16Short; break; 1841 case 4: STK = SimpleTypeKind::Int32; break; 1842 case 8: STK = SimpleTypeKind::Int64Quad; break; 1843 case 16: STK = SimpleTypeKind::Int128Oct; break; 1844 } 1845 break; 1846 case dwarf::DW_ATE_unsigned: 1847 switch (ByteSize) { 1848 case 1: STK = SimpleTypeKind::UnsignedCharacter; break; 1849 case 2: STK = SimpleTypeKind::UInt16Short; break; 1850 case 4: STK = SimpleTypeKind::UInt32; break; 1851 case 8: STK = SimpleTypeKind::UInt64Quad; break; 1852 case 16: STK = SimpleTypeKind::UInt128Oct; break; 1853 } 1854 break; 1855 case dwarf::DW_ATE_UTF: 1856 switch (ByteSize) { 1857 case 1: STK = SimpleTypeKind::Character8; break; 1858 case 2: STK = SimpleTypeKind::Character16; break; 1859 case 4: STK = SimpleTypeKind::Character32; break; 1860 } 1861 break; 1862 case dwarf::DW_ATE_signed_char: 1863 if (ByteSize == 1) 1864 STK = SimpleTypeKind::SignedCharacter; 1865 break; 1866 case dwarf::DW_ATE_unsigned_char: 1867 if (ByteSize == 1) 1868 STK = SimpleTypeKind::UnsignedCharacter; 1869 break; 1870 default: 1871 break; 1872 } 1873 1874 // Apply some fixups based on the source-level type name. 1875 // Include some amount of canonicalization from an old naming scheme Clang 1876 // used to use for integer types (in an outdated effort to be compatible with 1877 // GCC's debug info/GDB's behavior, which has since been addressed). 1878 if (STK == SimpleTypeKind::Int32 && 1879 (Ty->getName() == "long int" || Ty->getName() == "long")) 1880 STK = SimpleTypeKind::Int32Long; 1881 if (STK == SimpleTypeKind::UInt32 && (Ty->getName() == "long unsigned int" || 1882 Ty->getName() == "unsigned long")) 1883 STK = SimpleTypeKind::UInt32Long; 1884 if (STK == SimpleTypeKind::UInt16Short && 1885 (Ty->getName() == "wchar_t" || Ty->getName() == "__wchar_t")) 1886 STK = SimpleTypeKind::WideCharacter; 1887 if ((STK == SimpleTypeKind::SignedCharacter || 1888 STK == SimpleTypeKind::UnsignedCharacter) && 1889 Ty->getName() == "char") 1890 STK = SimpleTypeKind::NarrowCharacter; 1891 1892 return TypeIndex(STK); 1893 } 1894 1895 TypeIndex CodeViewDebug::lowerTypePointer(const DIDerivedType *Ty, 1896 PointerOptions PO) { 1897 TypeIndex PointeeTI = getTypeIndex(Ty->getBaseType()); 1898 1899 // Pointers to simple types without any options can use SimpleTypeMode, rather 1900 // than having a dedicated pointer type record. 1901 if (PointeeTI.isSimple() && PO == PointerOptions::None && 1902 PointeeTI.getSimpleMode() == SimpleTypeMode::Direct && 1903 Ty->getTag() == dwarf::DW_TAG_pointer_type) { 1904 SimpleTypeMode Mode = Ty->getSizeInBits() == 64 1905 ? SimpleTypeMode::NearPointer64 1906 : SimpleTypeMode::NearPointer32; 1907 return TypeIndex(PointeeTI.getSimpleKind(), Mode); 1908 } 1909 1910 PointerKind PK = 1911 Ty->getSizeInBits() == 64 ? PointerKind::Near64 : PointerKind::Near32; 1912 PointerMode PM = PointerMode::Pointer; 1913 switch (Ty->getTag()) { 1914 default: llvm_unreachable("not a pointer tag type"); 1915 case dwarf::DW_TAG_pointer_type: 1916 PM = PointerMode::Pointer; 1917 break; 1918 case dwarf::DW_TAG_reference_type: 1919 PM = PointerMode::LValueReference; 1920 break; 1921 case dwarf::DW_TAG_rvalue_reference_type: 1922 PM = PointerMode::RValueReference; 1923 break; 1924 } 1925 1926 if (Ty->isObjectPointer()) 1927 PO |= PointerOptions::Const; 1928 1929 PointerRecord PR(PointeeTI, PK, PM, PO, Ty->getSizeInBits() / 8); 1930 return TypeTable.writeLeafType(PR); 1931 } 1932 1933 static PointerToMemberRepresentation 1934 translatePtrToMemberRep(unsigned SizeInBytes, bool IsPMF, unsigned Flags) { 1935 // SizeInBytes being zero generally implies that the member pointer type was 1936 // incomplete, which can happen if it is part of a function prototype. In this 1937 // case, use the unknown model instead of the general model. 1938 if (IsPMF) { 1939 switch (Flags & DINode::FlagPtrToMemberRep) { 1940 case 0: 1941 return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown 1942 : PointerToMemberRepresentation::GeneralFunction; 1943 case DINode::FlagSingleInheritance: 1944 return PointerToMemberRepresentation::SingleInheritanceFunction; 1945 case DINode::FlagMultipleInheritance: 1946 return PointerToMemberRepresentation::MultipleInheritanceFunction; 1947 case DINode::FlagVirtualInheritance: 1948 return PointerToMemberRepresentation::VirtualInheritanceFunction; 1949 } 1950 } else { 1951 switch (Flags & DINode::FlagPtrToMemberRep) { 1952 case 0: 1953 return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown 1954 : PointerToMemberRepresentation::GeneralData; 1955 case DINode::FlagSingleInheritance: 1956 return PointerToMemberRepresentation::SingleInheritanceData; 1957 case DINode::FlagMultipleInheritance: 1958 return PointerToMemberRepresentation::MultipleInheritanceData; 1959 case DINode::FlagVirtualInheritance: 1960 return PointerToMemberRepresentation::VirtualInheritanceData; 1961 } 1962 } 1963 llvm_unreachable("invalid ptr to member representation"); 1964 } 1965 1966 TypeIndex CodeViewDebug::lowerTypeMemberPointer(const DIDerivedType *Ty, 1967 PointerOptions PO) { 1968 assert(Ty->getTag() == dwarf::DW_TAG_ptr_to_member_type); 1969 bool IsPMF = isa<DISubroutineType>(Ty->getBaseType()); 1970 TypeIndex ClassTI = getTypeIndex(Ty->getClassType()); 1971 TypeIndex PointeeTI = 1972 getTypeIndex(Ty->getBaseType(), IsPMF ? Ty->getClassType() : nullptr); 1973 PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64 1974 : PointerKind::Near32; 1975 PointerMode PM = IsPMF ? PointerMode::PointerToMemberFunction 1976 : PointerMode::PointerToDataMember; 1977 1978 assert(Ty->getSizeInBits() / 8 <= 0xff && "pointer size too big"); 1979 uint8_t SizeInBytes = Ty->getSizeInBits() / 8; 1980 MemberPointerInfo MPI( 1981 ClassTI, translatePtrToMemberRep(SizeInBytes, IsPMF, Ty->getFlags())); 1982 PointerRecord PR(PointeeTI, PK, PM, PO, SizeInBytes, MPI); 1983 return TypeTable.writeLeafType(PR); 1984 } 1985 1986 /// Given a DWARF calling convention, get the CodeView equivalent. If we don't 1987 /// have a translation, use the NearC convention. 1988 static CallingConvention dwarfCCToCodeView(unsigned DwarfCC) { 1989 switch (DwarfCC) { 1990 case dwarf::DW_CC_normal: return CallingConvention::NearC; 1991 case dwarf::DW_CC_BORLAND_msfastcall: return CallingConvention::NearFast; 1992 case dwarf::DW_CC_BORLAND_thiscall: return CallingConvention::ThisCall; 1993 case dwarf::DW_CC_BORLAND_stdcall: return CallingConvention::NearStdCall; 1994 case dwarf::DW_CC_BORLAND_pascal: return CallingConvention::NearPascal; 1995 case dwarf::DW_CC_LLVM_vectorcall: return CallingConvention::NearVector; 1996 } 1997 return CallingConvention::NearC; 1998 } 1999 2000 TypeIndex CodeViewDebug::lowerTypeModifier(const DIDerivedType *Ty) { 2001 ModifierOptions Mods = ModifierOptions::None; 2002 PointerOptions PO = PointerOptions::None; 2003 bool IsModifier = true; 2004 const DIType *BaseTy = Ty; 2005 while (IsModifier && BaseTy) { 2006 // FIXME: Need to add DWARF tags for __unaligned and _Atomic 2007 switch (BaseTy->getTag()) { 2008 case dwarf::DW_TAG_const_type: 2009 Mods |= ModifierOptions::Const; 2010 PO |= PointerOptions::Const; 2011 break; 2012 case dwarf::DW_TAG_volatile_type: 2013 Mods |= ModifierOptions::Volatile; 2014 PO |= PointerOptions::Volatile; 2015 break; 2016 case dwarf::DW_TAG_restrict_type: 2017 // Only pointer types be marked with __restrict. There is no known flag 2018 // for __restrict in LF_MODIFIER records. 2019 PO |= PointerOptions::Restrict; 2020 break; 2021 default: 2022 IsModifier = false; 2023 break; 2024 } 2025 if (IsModifier) 2026 BaseTy = cast<DIDerivedType>(BaseTy)->getBaseType(); 2027 } 2028 2029 // Check if the inner type will use an LF_POINTER record. If so, the 2030 // qualifiers will go in the LF_POINTER record. This comes up for types like 2031 // 'int *const' and 'int *__restrict', not the more common cases like 'const 2032 // char *'. 2033 if (BaseTy) { 2034 switch (BaseTy->getTag()) { 2035 case dwarf::DW_TAG_pointer_type: 2036 case dwarf::DW_TAG_reference_type: 2037 case dwarf::DW_TAG_rvalue_reference_type: 2038 return lowerTypePointer(cast<DIDerivedType>(BaseTy), PO); 2039 case dwarf::DW_TAG_ptr_to_member_type: 2040 return lowerTypeMemberPointer(cast<DIDerivedType>(BaseTy), PO); 2041 default: 2042 break; 2043 } 2044 } 2045 2046 TypeIndex ModifiedTI = getTypeIndex(BaseTy); 2047 2048 // Return the base type index if there aren't any modifiers. For example, the 2049 // metadata could contain restrict wrappers around non-pointer types. 2050 if (Mods == ModifierOptions::None) 2051 return ModifiedTI; 2052 2053 ModifierRecord MR(ModifiedTI, Mods); 2054 return TypeTable.writeLeafType(MR); 2055 } 2056 2057 TypeIndex CodeViewDebug::lowerTypeFunction(const DISubroutineType *Ty) { 2058 SmallVector<TypeIndex, 8> ReturnAndArgTypeIndices; 2059 for (const DIType *ArgType : Ty->getTypeArray()) 2060 ReturnAndArgTypeIndices.push_back(getTypeIndex(ArgType)); 2061 2062 // MSVC uses type none for variadic argument. 2063 if (ReturnAndArgTypeIndices.size() > 1 && 2064 ReturnAndArgTypeIndices.back() == TypeIndex::Void()) { 2065 ReturnAndArgTypeIndices.back() = TypeIndex::None(); 2066 } 2067 TypeIndex ReturnTypeIndex = TypeIndex::Void(); 2068 ArrayRef<TypeIndex> ArgTypeIndices = std::nullopt; 2069 if (!ReturnAndArgTypeIndices.empty()) { 2070 auto ReturnAndArgTypesRef = ArrayRef(ReturnAndArgTypeIndices); 2071 ReturnTypeIndex = ReturnAndArgTypesRef.front(); 2072 ArgTypeIndices = ReturnAndArgTypesRef.drop_front(); 2073 } 2074 2075 ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices); 2076 TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec); 2077 2078 CallingConvention CC = dwarfCCToCodeView(Ty->getCC()); 2079 2080 FunctionOptions FO = getFunctionOptions(Ty); 2081 ProcedureRecord Procedure(ReturnTypeIndex, CC, FO, ArgTypeIndices.size(), 2082 ArgListIndex); 2083 return TypeTable.writeLeafType(Procedure); 2084 } 2085 2086 TypeIndex CodeViewDebug::lowerTypeMemberFunction(const DISubroutineType *Ty, 2087 const DIType *ClassTy, 2088 int ThisAdjustment, 2089 bool IsStaticMethod, 2090 FunctionOptions FO) { 2091 // Lower the containing class type. 2092 TypeIndex ClassType = getTypeIndex(ClassTy); 2093 2094 DITypeRefArray ReturnAndArgs = Ty->getTypeArray(); 2095 2096 unsigned Index = 0; 2097 SmallVector<TypeIndex, 8> ArgTypeIndices; 2098 TypeIndex ReturnTypeIndex = TypeIndex::Void(); 2099 if (ReturnAndArgs.size() > Index) { 2100 ReturnTypeIndex = getTypeIndex(ReturnAndArgs[Index++]); 2101 } 2102 2103 // If the first argument is a pointer type and this isn't a static method, 2104 // treat it as the special 'this' parameter, which is encoded separately from 2105 // the arguments. 2106 TypeIndex ThisTypeIndex; 2107 if (!IsStaticMethod && ReturnAndArgs.size() > Index) { 2108 if (const DIDerivedType *PtrTy = 2109 dyn_cast_or_null<DIDerivedType>(ReturnAndArgs[Index])) { 2110 if (PtrTy->getTag() == dwarf::DW_TAG_pointer_type) { 2111 ThisTypeIndex = getTypeIndexForThisPtr(PtrTy, Ty); 2112 Index++; 2113 } 2114 } 2115 } 2116 2117 while (Index < ReturnAndArgs.size()) 2118 ArgTypeIndices.push_back(getTypeIndex(ReturnAndArgs[Index++])); 2119 2120 // MSVC uses type none for variadic argument. 2121 if (!ArgTypeIndices.empty() && ArgTypeIndices.back() == TypeIndex::Void()) 2122 ArgTypeIndices.back() = TypeIndex::None(); 2123 2124 ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices); 2125 TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec); 2126 2127 CallingConvention CC = dwarfCCToCodeView(Ty->getCC()); 2128 2129 MemberFunctionRecord MFR(ReturnTypeIndex, ClassType, ThisTypeIndex, CC, FO, 2130 ArgTypeIndices.size(), ArgListIndex, ThisAdjustment); 2131 return TypeTable.writeLeafType(MFR); 2132 } 2133 2134 TypeIndex CodeViewDebug::lowerTypeVFTableShape(const DIDerivedType *Ty) { 2135 unsigned VSlotCount = 2136 Ty->getSizeInBits() / (8 * Asm->MAI->getCodePointerSize()); 2137 SmallVector<VFTableSlotKind, 4> Slots(VSlotCount, VFTableSlotKind::Near); 2138 2139 VFTableShapeRecord VFTSR(Slots); 2140 return TypeTable.writeLeafType(VFTSR); 2141 } 2142 2143 static MemberAccess translateAccessFlags(unsigned RecordTag, unsigned Flags) { 2144 switch (Flags & DINode::FlagAccessibility) { 2145 case DINode::FlagPrivate: return MemberAccess::Private; 2146 case DINode::FlagPublic: return MemberAccess::Public; 2147 case DINode::FlagProtected: return MemberAccess::Protected; 2148 case 0: 2149 // If there was no explicit access control, provide the default for the tag. 2150 return RecordTag == dwarf::DW_TAG_class_type ? MemberAccess::Private 2151 : MemberAccess::Public; 2152 } 2153 llvm_unreachable("access flags are exclusive"); 2154 } 2155 2156 static MethodOptions translateMethodOptionFlags(const DISubprogram *SP) { 2157 if (SP->isArtificial()) 2158 return MethodOptions::CompilerGenerated; 2159 2160 // FIXME: Handle other MethodOptions. 2161 2162 return MethodOptions::None; 2163 } 2164 2165 static MethodKind translateMethodKindFlags(const DISubprogram *SP, 2166 bool Introduced) { 2167 if (SP->getFlags() & DINode::FlagStaticMember) 2168 return MethodKind::Static; 2169 2170 switch (SP->getVirtuality()) { 2171 case dwarf::DW_VIRTUALITY_none: 2172 break; 2173 case dwarf::DW_VIRTUALITY_virtual: 2174 return Introduced ? MethodKind::IntroducingVirtual : MethodKind::Virtual; 2175 case dwarf::DW_VIRTUALITY_pure_virtual: 2176 return Introduced ? MethodKind::PureIntroducingVirtual 2177 : MethodKind::PureVirtual; 2178 default: 2179 llvm_unreachable("unhandled virtuality case"); 2180 } 2181 2182 return MethodKind::Vanilla; 2183 } 2184 2185 static TypeRecordKind getRecordKind(const DICompositeType *Ty) { 2186 switch (Ty->getTag()) { 2187 case dwarf::DW_TAG_class_type: 2188 return TypeRecordKind::Class; 2189 case dwarf::DW_TAG_structure_type: 2190 return TypeRecordKind::Struct; 2191 default: 2192 llvm_unreachable("unexpected tag"); 2193 } 2194 } 2195 2196 /// Return ClassOptions that should be present on both the forward declaration 2197 /// and the defintion of a tag type. 2198 static ClassOptions getCommonClassOptions(const DICompositeType *Ty) { 2199 ClassOptions CO = ClassOptions::None; 2200 2201 // MSVC always sets this flag, even for local types. Clang doesn't always 2202 // appear to give every type a linkage name, which may be problematic for us. 2203 // FIXME: Investigate the consequences of not following them here. 2204 if (!Ty->getIdentifier().empty()) 2205 CO |= ClassOptions::HasUniqueName; 2206 2207 // Put the Nested flag on a type if it appears immediately inside a tag type. 2208 // Do not walk the scope chain. Do not attempt to compute ContainsNestedClass 2209 // here. That flag is only set on definitions, and not forward declarations. 2210 const DIScope *ImmediateScope = Ty->getScope(); 2211 if (ImmediateScope && isa<DICompositeType>(ImmediateScope)) 2212 CO |= ClassOptions::Nested; 2213 2214 // Put the Scoped flag on function-local types. MSVC puts this flag for enum 2215 // type only when it has an immediate function scope. Clang never puts enums 2216 // inside DILexicalBlock scopes. Enum types, as generated by clang, are 2217 // always in function, class, or file scopes. 2218 if (Ty->getTag() == dwarf::DW_TAG_enumeration_type) { 2219 if (ImmediateScope && isa<DISubprogram>(ImmediateScope)) 2220 CO |= ClassOptions::Scoped; 2221 } else { 2222 for (const DIScope *Scope = ImmediateScope; Scope != nullptr; 2223 Scope = Scope->getScope()) { 2224 if (isa<DISubprogram>(Scope)) { 2225 CO |= ClassOptions::Scoped; 2226 break; 2227 } 2228 } 2229 } 2230 2231 return CO; 2232 } 2233 2234 void CodeViewDebug::addUDTSrcLine(const DIType *Ty, TypeIndex TI) { 2235 switch (Ty->getTag()) { 2236 case dwarf::DW_TAG_class_type: 2237 case dwarf::DW_TAG_structure_type: 2238 case dwarf::DW_TAG_union_type: 2239 case dwarf::DW_TAG_enumeration_type: 2240 break; 2241 default: 2242 return; 2243 } 2244 2245 if (const auto *File = Ty->getFile()) { 2246 StringIdRecord SIDR(TypeIndex(0x0), getFullFilepath(File)); 2247 TypeIndex SIDI = TypeTable.writeLeafType(SIDR); 2248 2249 UdtSourceLineRecord USLR(TI, SIDI, Ty->getLine()); 2250 TypeTable.writeLeafType(USLR); 2251 } 2252 } 2253 2254 TypeIndex CodeViewDebug::lowerTypeEnum(const DICompositeType *Ty) { 2255 ClassOptions CO = getCommonClassOptions(Ty); 2256 TypeIndex FTI; 2257 unsigned EnumeratorCount = 0; 2258 2259 if (Ty->isForwardDecl()) { 2260 CO |= ClassOptions::ForwardReference; 2261 } else { 2262 ContinuationRecordBuilder ContinuationBuilder; 2263 ContinuationBuilder.begin(ContinuationRecordKind::FieldList); 2264 for (const DINode *Element : Ty->getElements()) { 2265 // We assume that the frontend provides all members in source declaration 2266 // order, which is what MSVC does. 2267 if (auto *Enumerator = dyn_cast_or_null<DIEnumerator>(Element)) { 2268 // FIXME: Is it correct to always emit these as unsigned here? 2269 EnumeratorRecord ER(MemberAccess::Public, 2270 APSInt(Enumerator->getValue(), true), 2271 Enumerator->getName()); 2272 ContinuationBuilder.writeMemberType(ER); 2273 EnumeratorCount++; 2274 } 2275 } 2276 FTI = TypeTable.insertRecord(ContinuationBuilder); 2277 } 2278 2279 std::string FullName = getFullyQualifiedName(Ty); 2280 2281 EnumRecord ER(EnumeratorCount, CO, FTI, FullName, Ty->getIdentifier(), 2282 getTypeIndex(Ty->getBaseType())); 2283 TypeIndex EnumTI = TypeTable.writeLeafType(ER); 2284 2285 addUDTSrcLine(Ty, EnumTI); 2286 2287 return EnumTI; 2288 } 2289 2290 //===----------------------------------------------------------------------===// 2291 // ClassInfo 2292 //===----------------------------------------------------------------------===// 2293 2294 struct llvm::ClassInfo { 2295 struct MemberInfo { 2296 const DIDerivedType *MemberTypeNode; 2297 uint64_t BaseOffset; 2298 }; 2299 // [MemberInfo] 2300 using MemberList = std::vector<MemberInfo>; 2301 2302 using MethodsList = TinyPtrVector<const DISubprogram *>; 2303 // MethodName -> MethodsList 2304 using MethodsMap = MapVector<MDString *, MethodsList>; 2305 2306 /// Base classes. 2307 std::vector<const DIDerivedType *> Inheritance; 2308 2309 /// Direct members. 2310 MemberList Members; 2311 // Direct overloaded methods gathered by name. 2312 MethodsMap Methods; 2313 2314 TypeIndex VShapeTI; 2315 2316 std::vector<const DIType *> NestedTypes; 2317 }; 2318 2319 void CodeViewDebug::clear() { 2320 assert(CurFn == nullptr); 2321 FileIdMap.clear(); 2322 FnDebugInfo.clear(); 2323 FileToFilepathMap.clear(); 2324 LocalUDTs.clear(); 2325 GlobalUDTs.clear(); 2326 TypeIndices.clear(); 2327 CompleteTypeIndices.clear(); 2328 ScopeGlobals.clear(); 2329 CVGlobalVariableOffsets.clear(); 2330 } 2331 2332 void CodeViewDebug::collectMemberInfo(ClassInfo &Info, 2333 const DIDerivedType *DDTy) { 2334 if (!DDTy->getName().empty()) { 2335 Info.Members.push_back({DDTy, 0}); 2336 2337 // Collect static const data members with values. 2338 if ((DDTy->getFlags() & DINode::FlagStaticMember) == 2339 DINode::FlagStaticMember) { 2340 if (DDTy->getConstant() && (isa<ConstantInt>(DDTy->getConstant()) || 2341 isa<ConstantFP>(DDTy->getConstant()))) 2342 StaticConstMembers.push_back(DDTy); 2343 } 2344 2345 return; 2346 } 2347 2348 // An unnamed member may represent a nested struct or union. Attempt to 2349 // interpret the unnamed member as a DICompositeType possibly wrapped in 2350 // qualifier types. Add all the indirect fields to the current record if that 2351 // succeeds, and drop the member if that fails. 2352 assert((DDTy->getOffsetInBits() % 8) == 0 && "Unnamed bitfield member!"); 2353 uint64_t Offset = DDTy->getOffsetInBits(); 2354 const DIType *Ty = DDTy->getBaseType(); 2355 bool FullyResolved = false; 2356 while (!FullyResolved) { 2357 switch (Ty->getTag()) { 2358 case dwarf::DW_TAG_const_type: 2359 case dwarf::DW_TAG_volatile_type: 2360 // FIXME: we should apply the qualifier types to the indirect fields 2361 // rather than dropping them. 2362 Ty = cast<DIDerivedType>(Ty)->getBaseType(); 2363 break; 2364 default: 2365 FullyResolved = true; 2366 break; 2367 } 2368 } 2369 2370 const DICompositeType *DCTy = dyn_cast<DICompositeType>(Ty); 2371 if (!DCTy) 2372 return; 2373 2374 ClassInfo NestedInfo = collectClassInfo(DCTy); 2375 for (const ClassInfo::MemberInfo &IndirectField : NestedInfo.Members) 2376 Info.Members.push_back( 2377 {IndirectField.MemberTypeNode, IndirectField.BaseOffset + Offset}); 2378 } 2379 2380 ClassInfo CodeViewDebug::collectClassInfo(const DICompositeType *Ty) { 2381 ClassInfo Info; 2382 // Add elements to structure type. 2383 DINodeArray Elements = Ty->getElements(); 2384 for (auto *Element : Elements) { 2385 // We assume that the frontend provides all members in source declaration 2386 // order, which is what MSVC does. 2387 if (!Element) 2388 continue; 2389 if (auto *SP = dyn_cast<DISubprogram>(Element)) { 2390 Info.Methods[SP->getRawName()].push_back(SP); 2391 } else if (auto *DDTy = dyn_cast<DIDerivedType>(Element)) { 2392 if (DDTy->getTag() == dwarf::DW_TAG_member) { 2393 collectMemberInfo(Info, DDTy); 2394 } else if (DDTy->getTag() == dwarf::DW_TAG_inheritance) { 2395 Info.Inheritance.push_back(DDTy); 2396 } else if (DDTy->getTag() == dwarf::DW_TAG_pointer_type && 2397 DDTy->getName() == "__vtbl_ptr_type") { 2398 Info.VShapeTI = getTypeIndex(DDTy); 2399 } else if (DDTy->getTag() == dwarf::DW_TAG_typedef) { 2400 Info.NestedTypes.push_back(DDTy); 2401 } else if (DDTy->getTag() == dwarf::DW_TAG_friend) { 2402 // Ignore friend members. It appears that MSVC emitted info about 2403 // friends in the past, but modern versions do not. 2404 } 2405 } else if (auto *Composite = dyn_cast<DICompositeType>(Element)) { 2406 Info.NestedTypes.push_back(Composite); 2407 } 2408 // Skip other unrecognized kinds of elements. 2409 } 2410 return Info; 2411 } 2412 2413 static bool shouldAlwaysEmitCompleteClassType(const DICompositeType *Ty) { 2414 // This routine is used by lowerTypeClass and lowerTypeUnion to determine 2415 // if a complete type should be emitted instead of a forward reference. 2416 return Ty->getName().empty() && Ty->getIdentifier().empty() && 2417 !Ty->isForwardDecl(); 2418 } 2419 2420 TypeIndex CodeViewDebug::lowerTypeClass(const DICompositeType *Ty) { 2421 // Emit the complete type for unnamed structs. C++ classes with methods 2422 // which have a circular reference back to the class type are expected to 2423 // be named by the front-end and should not be "unnamed". C unnamed 2424 // structs should not have circular references. 2425 if (shouldAlwaysEmitCompleteClassType(Ty)) { 2426 // If this unnamed complete type is already in the process of being defined 2427 // then the description of the type is malformed and cannot be emitted 2428 // into CodeView correctly so report a fatal error. 2429 auto I = CompleteTypeIndices.find(Ty); 2430 if (I != CompleteTypeIndices.end() && I->second == TypeIndex()) 2431 report_fatal_error("cannot debug circular reference to unnamed type"); 2432 return getCompleteTypeIndex(Ty); 2433 } 2434 2435 // First, construct the forward decl. Don't look into Ty to compute the 2436 // forward decl options, since it might not be available in all TUs. 2437 TypeRecordKind Kind = getRecordKind(Ty); 2438 ClassOptions CO = 2439 ClassOptions::ForwardReference | getCommonClassOptions(Ty); 2440 std::string FullName = getFullyQualifiedName(Ty); 2441 ClassRecord CR(Kind, 0, CO, TypeIndex(), TypeIndex(), TypeIndex(), 0, 2442 FullName, Ty->getIdentifier()); 2443 TypeIndex FwdDeclTI = TypeTable.writeLeafType(CR); 2444 if (!Ty->isForwardDecl()) 2445 DeferredCompleteTypes.push_back(Ty); 2446 return FwdDeclTI; 2447 } 2448 2449 TypeIndex CodeViewDebug::lowerCompleteTypeClass(const DICompositeType *Ty) { 2450 // Construct the field list and complete type record. 2451 TypeRecordKind Kind = getRecordKind(Ty); 2452 ClassOptions CO = getCommonClassOptions(Ty); 2453 TypeIndex FieldTI; 2454 TypeIndex VShapeTI; 2455 unsigned FieldCount; 2456 bool ContainsNestedClass; 2457 std::tie(FieldTI, VShapeTI, FieldCount, ContainsNestedClass) = 2458 lowerRecordFieldList(Ty); 2459 2460 if (ContainsNestedClass) 2461 CO |= ClassOptions::ContainsNestedClass; 2462 2463 // MSVC appears to set this flag by searching any destructor or method with 2464 // FunctionOptions::Constructor among the emitted members. Clang AST has all 2465 // the members, however special member functions are not yet emitted into 2466 // debug information. For now checking a class's non-triviality seems enough. 2467 // FIXME: not true for a nested unnamed struct. 2468 if (isNonTrivial(Ty)) 2469 CO |= ClassOptions::HasConstructorOrDestructor; 2470 2471 std::string FullName = getFullyQualifiedName(Ty); 2472 2473 uint64_t SizeInBytes = Ty->getSizeInBits() / 8; 2474 2475 ClassRecord CR(Kind, FieldCount, CO, FieldTI, TypeIndex(), VShapeTI, 2476 SizeInBytes, FullName, Ty->getIdentifier()); 2477 TypeIndex ClassTI = TypeTable.writeLeafType(CR); 2478 2479 addUDTSrcLine(Ty, ClassTI); 2480 2481 addToUDTs(Ty); 2482 2483 return ClassTI; 2484 } 2485 2486 TypeIndex CodeViewDebug::lowerTypeUnion(const DICompositeType *Ty) { 2487 // Emit the complete type for unnamed unions. 2488 if (shouldAlwaysEmitCompleteClassType(Ty)) 2489 return getCompleteTypeIndex(Ty); 2490 2491 ClassOptions CO = 2492 ClassOptions::ForwardReference | getCommonClassOptions(Ty); 2493 std::string FullName = getFullyQualifiedName(Ty); 2494 UnionRecord UR(0, CO, TypeIndex(), 0, FullName, Ty->getIdentifier()); 2495 TypeIndex FwdDeclTI = TypeTable.writeLeafType(UR); 2496 if (!Ty->isForwardDecl()) 2497 DeferredCompleteTypes.push_back(Ty); 2498 return FwdDeclTI; 2499 } 2500 2501 TypeIndex CodeViewDebug::lowerCompleteTypeUnion(const DICompositeType *Ty) { 2502 ClassOptions CO = ClassOptions::Sealed | getCommonClassOptions(Ty); 2503 TypeIndex FieldTI; 2504 unsigned FieldCount; 2505 bool ContainsNestedClass; 2506 std::tie(FieldTI, std::ignore, FieldCount, ContainsNestedClass) = 2507 lowerRecordFieldList(Ty); 2508 2509 if (ContainsNestedClass) 2510 CO |= ClassOptions::ContainsNestedClass; 2511 2512 uint64_t SizeInBytes = Ty->getSizeInBits() / 8; 2513 std::string FullName = getFullyQualifiedName(Ty); 2514 2515 UnionRecord UR(FieldCount, CO, FieldTI, SizeInBytes, FullName, 2516 Ty->getIdentifier()); 2517 TypeIndex UnionTI = TypeTable.writeLeafType(UR); 2518 2519 addUDTSrcLine(Ty, UnionTI); 2520 2521 addToUDTs(Ty); 2522 2523 return UnionTI; 2524 } 2525 2526 std::tuple<TypeIndex, TypeIndex, unsigned, bool> 2527 CodeViewDebug::lowerRecordFieldList(const DICompositeType *Ty) { 2528 // Manually count members. MSVC appears to count everything that generates a 2529 // field list record. Each individual overload in a method overload group 2530 // contributes to this count, even though the overload group is a single field 2531 // list record. 2532 unsigned MemberCount = 0; 2533 ClassInfo Info = collectClassInfo(Ty); 2534 ContinuationRecordBuilder ContinuationBuilder; 2535 ContinuationBuilder.begin(ContinuationRecordKind::FieldList); 2536 2537 // Create base classes. 2538 for (const DIDerivedType *I : Info.Inheritance) { 2539 if (I->getFlags() & DINode::FlagVirtual) { 2540 // Virtual base. 2541 unsigned VBPtrOffset = I->getVBPtrOffset(); 2542 // FIXME: Despite the accessor name, the offset is really in bytes. 2543 unsigned VBTableIndex = I->getOffsetInBits() / 4; 2544 auto RecordKind = (I->getFlags() & DINode::FlagIndirectVirtualBase) == DINode::FlagIndirectVirtualBase 2545 ? TypeRecordKind::IndirectVirtualBaseClass 2546 : TypeRecordKind::VirtualBaseClass; 2547 VirtualBaseClassRecord VBCR( 2548 RecordKind, translateAccessFlags(Ty->getTag(), I->getFlags()), 2549 getTypeIndex(I->getBaseType()), getVBPTypeIndex(), VBPtrOffset, 2550 VBTableIndex); 2551 2552 ContinuationBuilder.writeMemberType(VBCR); 2553 MemberCount++; 2554 } else { 2555 assert(I->getOffsetInBits() % 8 == 0 && 2556 "bases must be on byte boundaries"); 2557 BaseClassRecord BCR(translateAccessFlags(Ty->getTag(), I->getFlags()), 2558 getTypeIndex(I->getBaseType()), 2559 I->getOffsetInBits() / 8); 2560 ContinuationBuilder.writeMemberType(BCR); 2561 MemberCount++; 2562 } 2563 } 2564 2565 // Create members. 2566 for (ClassInfo::MemberInfo &MemberInfo : Info.Members) { 2567 const DIDerivedType *Member = MemberInfo.MemberTypeNode; 2568 TypeIndex MemberBaseType = getTypeIndex(Member->getBaseType()); 2569 StringRef MemberName = Member->getName(); 2570 MemberAccess Access = 2571 translateAccessFlags(Ty->getTag(), Member->getFlags()); 2572 2573 if (Member->isStaticMember()) { 2574 StaticDataMemberRecord SDMR(Access, MemberBaseType, MemberName); 2575 ContinuationBuilder.writeMemberType(SDMR); 2576 MemberCount++; 2577 continue; 2578 } 2579 2580 // Virtual function pointer member. 2581 if ((Member->getFlags() & DINode::FlagArtificial) && 2582 Member->getName().startswith("_vptr$")) { 2583 VFPtrRecord VFPR(getTypeIndex(Member->getBaseType())); 2584 ContinuationBuilder.writeMemberType(VFPR); 2585 MemberCount++; 2586 continue; 2587 } 2588 2589 // Data member. 2590 uint64_t MemberOffsetInBits = 2591 Member->getOffsetInBits() + MemberInfo.BaseOffset; 2592 if (Member->isBitField()) { 2593 uint64_t StartBitOffset = MemberOffsetInBits; 2594 if (const auto *CI = 2595 dyn_cast_or_null<ConstantInt>(Member->getStorageOffsetInBits())) { 2596 MemberOffsetInBits = CI->getZExtValue() + MemberInfo.BaseOffset; 2597 } 2598 StartBitOffset -= MemberOffsetInBits; 2599 BitFieldRecord BFR(MemberBaseType, Member->getSizeInBits(), 2600 StartBitOffset); 2601 MemberBaseType = TypeTable.writeLeafType(BFR); 2602 } 2603 uint64_t MemberOffsetInBytes = MemberOffsetInBits / 8; 2604 DataMemberRecord DMR(Access, MemberBaseType, MemberOffsetInBytes, 2605 MemberName); 2606 ContinuationBuilder.writeMemberType(DMR); 2607 MemberCount++; 2608 } 2609 2610 // Create methods 2611 for (auto &MethodItr : Info.Methods) { 2612 StringRef Name = MethodItr.first->getString(); 2613 2614 std::vector<OneMethodRecord> Methods; 2615 for (const DISubprogram *SP : MethodItr.second) { 2616 TypeIndex MethodType = getMemberFunctionType(SP, Ty); 2617 bool Introduced = SP->getFlags() & DINode::FlagIntroducedVirtual; 2618 2619 unsigned VFTableOffset = -1; 2620 if (Introduced) 2621 VFTableOffset = SP->getVirtualIndex() * getPointerSizeInBytes(); 2622 2623 Methods.push_back(OneMethodRecord( 2624 MethodType, translateAccessFlags(Ty->getTag(), SP->getFlags()), 2625 translateMethodKindFlags(SP, Introduced), 2626 translateMethodOptionFlags(SP), VFTableOffset, Name)); 2627 MemberCount++; 2628 } 2629 assert(!Methods.empty() && "Empty methods map entry"); 2630 if (Methods.size() == 1) 2631 ContinuationBuilder.writeMemberType(Methods[0]); 2632 else { 2633 // FIXME: Make this use its own ContinuationBuilder so that 2634 // MethodOverloadList can be split correctly. 2635 MethodOverloadListRecord MOLR(Methods); 2636 TypeIndex MethodList = TypeTable.writeLeafType(MOLR); 2637 2638 OverloadedMethodRecord OMR(Methods.size(), MethodList, Name); 2639 ContinuationBuilder.writeMemberType(OMR); 2640 } 2641 } 2642 2643 // Create nested classes. 2644 for (const DIType *Nested : Info.NestedTypes) { 2645 NestedTypeRecord R(getTypeIndex(Nested), Nested->getName()); 2646 ContinuationBuilder.writeMemberType(R); 2647 MemberCount++; 2648 } 2649 2650 TypeIndex FieldTI = TypeTable.insertRecord(ContinuationBuilder); 2651 return std::make_tuple(FieldTI, Info.VShapeTI, MemberCount, 2652 !Info.NestedTypes.empty()); 2653 } 2654 2655 TypeIndex CodeViewDebug::getVBPTypeIndex() { 2656 if (!VBPType.getIndex()) { 2657 // Make a 'const int *' type. 2658 ModifierRecord MR(TypeIndex::Int32(), ModifierOptions::Const); 2659 TypeIndex ModifiedTI = TypeTable.writeLeafType(MR); 2660 2661 PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64 2662 : PointerKind::Near32; 2663 PointerMode PM = PointerMode::Pointer; 2664 PointerOptions PO = PointerOptions::None; 2665 PointerRecord PR(ModifiedTI, PK, PM, PO, getPointerSizeInBytes()); 2666 VBPType = TypeTable.writeLeafType(PR); 2667 } 2668 2669 return VBPType; 2670 } 2671 2672 TypeIndex CodeViewDebug::getTypeIndex(const DIType *Ty, const DIType *ClassTy) { 2673 // The null DIType is the void type. Don't try to hash it. 2674 if (!Ty) 2675 return TypeIndex::Void(); 2676 2677 // Check if we've already translated this type. Don't try to do a 2678 // get-or-create style insertion that caches the hash lookup across the 2679 // lowerType call. It will update the TypeIndices map. 2680 auto I = TypeIndices.find({Ty, ClassTy}); 2681 if (I != TypeIndices.end()) 2682 return I->second; 2683 2684 TypeLoweringScope S(*this); 2685 TypeIndex TI = lowerType(Ty, ClassTy); 2686 return recordTypeIndexForDINode(Ty, TI, ClassTy); 2687 } 2688 2689 codeview::TypeIndex 2690 CodeViewDebug::getTypeIndexForThisPtr(const DIDerivedType *PtrTy, 2691 const DISubroutineType *SubroutineTy) { 2692 assert(PtrTy->getTag() == dwarf::DW_TAG_pointer_type && 2693 "this type must be a pointer type"); 2694 2695 PointerOptions Options = PointerOptions::None; 2696 if (SubroutineTy->getFlags() & DINode::DIFlags::FlagLValueReference) 2697 Options = PointerOptions::LValueRefThisPointer; 2698 else if (SubroutineTy->getFlags() & DINode::DIFlags::FlagRValueReference) 2699 Options = PointerOptions::RValueRefThisPointer; 2700 2701 // Check if we've already translated this type. If there is no ref qualifier 2702 // on the function then we look up this pointer type with no associated class 2703 // so that the TypeIndex for the this pointer can be shared with the type 2704 // index for other pointers to this class type. If there is a ref qualifier 2705 // then we lookup the pointer using the subroutine as the parent type. 2706 auto I = TypeIndices.find({PtrTy, SubroutineTy}); 2707 if (I != TypeIndices.end()) 2708 return I->second; 2709 2710 TypeLoweringScope S(*this); 2711 TypeIndex TI = lowerTypePointer(PtrTy, Options); 2712 return recordTypeIndexForDINode(PtrTy, TI, SubroutineTy); 2713 } 2714 2715 TypeIndex CodeViewDebug::getTypeIndexForReferenceTo(const DIType *Ty) { 2716 PointerRecord PR(getTypeIndex(Ty), 2717 getPointerSizeInBytes() == 8 ? PointerKind::Near64 2718 : PointerKind::Near32, 2719 PointerMode::LValueReference, PointerOptions::None, 2720 Ty->getSizeInBits() / 8); 2721 return TypeTable.writeLeafType(PR); 2722 } 2723 2724 TypeIndex CodeViewDebug::getCompleteTypeIndex(const DIType *Ty) { 2725 // The null DIType is the void type. Don't try to hash it. 2726 if (!Ty) 2727 return TypeIndex::Void(); 2728 2729 // Look through typedefs when getting the complete type index. Call 2730 // getTypeIndex on the typdef to ensure that any UDTs are accumulated and are 2731 // emitted only once. 2732 if (Ty->getTag() == dwarf::DW_TAG_typedef) 2733 (void)getTypeIndex(Ty); 2734 while (Ty->getTag() == dwarf::DW_TAG_typedef) 2735 Ty = cast<DIDerivedType>(Ty)->getBaseType(); 2736 2737 // If this is a non-record type, the complete type index is the same as the 2738 // normal type index. Just call getTypeIndex. 2739 switch (Ty->getTag()) { 2740 case dwarf::DW_TAG_class_type: 2741 case dwarf::DW_TAG_structure_type: 2742 case dwarf::DW_TAG_union_type: 2743 break; 2744 default: 2745 return getTypeIndex(Ty); 2746 } 2747 2748 const auto *CTy = cast<DICompositeType>(Ty); 2749 2750 TypeLoweringScope S(*this); 2751 2752 // Make sure the forward declaration is emitted first. It's unclear if this 2753 // is necessary, but MSVC does it, and we should follow suit until we can show 2754 // otherwise. 2755 // We only emit a forward declaration for named types. 2756 if (!CTy->getName().empty() || !CTy->getIdentifier().empty()) { 2757 TypeIndex FwdDeclTI = getTypeIndex(CTy); 2758 2759 // Just use the forward decl if we don't have complete type info. This 2760 // might happen if the frontend is using modules and expects the complete 2761 // definition to be emitted elsewhere. 2762 if (CTy->isForwardDecl()) 2763 return FwdDeclTI; 2764 } 2765 2766 // Check if we've already translated the complete record type. 2767 // Insert the type with a null TypeIndex to signify that the type is currently 2768 // being lowered. 2769 auto InsertResult = CompleteTypeIndices.insert({CTy, TypeIndex()}); 2770 if (!InsertResult.second) 2771 return InsertResult.first->second; 2772 2773 TypeIndex TI; 2774 switch (CTy->getTag()) { 2775 case dwarf::DW_TAG_class_type: 2776 case dwarf::DW_TAG_structure_type: 2777 TI = lowerCompleteTypeClass(CTy); 2778 break; 2779 case dwarf::DW_TAG_union_type: 2780 TI = lowerCompleteTypeUnion(CTy); 2781 break; 2782 default: 2783 llvm_unreachable("not a record"); 2784 } 2785 2786 // Update the type index associated with this CompositeType. This cannot 2787 // use the 'InsertResult' iterator above because it is potentially 2788 // invalidated by map insertions which can occur while lowering the class 2789 // type above. 2790 CompleteTypeIndices[CTy] = TI; 2791 return TI; 2792 } 2793 2794 /// Emit all the deferred complete record types. Try to do this in FIFO order, 2795 /// and do this until fixpoint, as each complete record type typically 2796 /// references 2797 /// many other record types. 2798 void CodeViewDebug::emitDeferredCompleteTypes() { 2799 SmallVector<const DICompositeType *, 4> TypesToEmit; 2800 while (!DeferredCompleteTypes.empty()) { 2801 std::swap(DeferredCompleteTypes, TypesToEmit); 2802 for (const DICompositeType *RecordTy : TypesToEmit) 2803 getCompleteTypeIndex(RecordTy); 2804 TypesToEmit.clear(); 2805 } 2806 } 2807 2808 void CodeViewDebug::emitLocalVariableList(const FunctionInfo &FI, 2809 ArrayRef<LocalVariable> Locals) { 2810 // Get the sorted list of parameters and emit them first. 2811 SmallVector<const LocalVariable *, 6> Params; 2812 for (const LocalVariable &L : Locals) 2813 if (L.DIVar->isParameter()) 2814 Params.push_back(&L); 2815 llvm::sort(Params, [](const LocalVariable *L, const LocalVariable *R) { 2816 return L->DIVar->getArg() < R->DIVar->getArg(); 2817 }); 2818 for (const LocalVariable *L : Params) 2819 emitLocalVariable(FI, *L); 2820 2821 // Next emit all non-parameters in the order that we found them. 2822 for (const LocalVariable &L : Locals) { 2823 if (!L.DIVar->isParameter()) { 2824 if (L.ConstantValue) { 2825 // If ConstantValue is set we will emit it as a S_CONSTANT instead of a 2826 // S_LOCAL in order to be able to represent it at all. 2827 const DIType *Ty = L.DIVar->getType(); 2828 APSInt Val(*L.ConstantValue); 2829 emitConstantSymbolRecord(Ty, Val, std::string(L.DIVar->getName())); 2830 } else { 2831 emitLocalVariable(FI, L); 2832 } 2833 } 2834 } 2835 } 2836 2837 void CodeViewDebug::emitLocalVariable(const FunctionInfo &FI, 2838 const LocalVariable &Var) { 2839 // LocalSym record, see SymbolRecord.h for more info. 2840 MCSymbol *LocalEnd = beginSymbolRecord(SymbolKind::S_LOCAL); 2841 2842 LocalSymFlags Flags = LocalSymFlags::None; 2843 if (Var.DIVar->isParameter()) 2844 Flags |= LocalSymFlags::IsParameter; 2845 if (Var.DefRanges.empty()) 2846 Flags |= LocalSymFlags::IsOptimizedOut; 2847 2848 OS.AddComment("TypeIndex"); 2849 TypeIndex TI = Var.UseReferenceType 2850 ? getTypeIndexForReferenceTo(Var.DIVar->getType()) 2851 : getCompleteTypeIndex(Var.DIVar->getType()); 2852 OS.emitInt32(TI.getIndex()); 2853 OS.AddComment("Flags"); 2854 OS.emitInt16(static_cast<uint16_t>(Flags)); 2855 // Truncate the name so we won't overflow the record length field. 2856 emitNullTerminatedSymbolName(OS, Var.DIVar->getName()); 2857 endSymbolRecord(LocalEnd); 2858 2859 // Calculate the on disk prefix of the appropriate def range record. The 2860 // records and on disk formats are described in SymbolRecords.h. BytePrefix 2861 // should be big enough to hold all forms without memory allocation. 2862 SmallString<20> BytePrefix; 2863 for (const auto &Pair : Var.DefRanges) { 2864 LocalVarDef DefRange = Pair.first; 2865 const auto &Ranges = Pair.second; 2866 BytePrefix.clear(); 2867 if (DefRange.InMemory) { 2868 int Offset = DefRange.DataOffset; 2869 unsigned Reg = DefRange.CVRegister; 2870 2871 // 32-bit x86 call sequences often use PUSH instructions, which disrupt 2872 // ESP-relative offsets. Use the virtual frame pointer, VFRAME or $T0, 2873 // instead. In frames without stack realignment, $T0 will be the CFA. 2874 if (RegisterId(Reg) == RegisterId::ESP) { 2875 Reg = unsigned(RegisterId::VFRAME); 2876 Offset += FI.OffsetAdjustment; 2877 } 2878 2879 // If we can use the chosen frame pointer for the frame and this isn't a 2880 // sliced aggregate, use the smaller S_DEFRANGE_FRAMEPOINTER_REL record. 2881 // Otherwise, use S_DEFRANGE_REGISTER_REL. 2882 EncodedFramePtrReg EncFP = encodeFramePtrReg(RegisterId(Reg), TheCPU); 2883 if (!DefRange.IsSubfield && EncFP != EncodedFramePtrReg::None && 2884 (bool(Flags & LocalSymFlags::IsParameter) 2885 ? (EncFP == FI.EncodedParamFramePtrReg) 2886 : (EncFP == FI.EncodedLocalFramePtrReg))) { 2887 DefRangeFramePointerRelHeader DRHdr; 2888 DRHdr.Offset = Offset; 2889 OS.emitCVDefRangeDirective(Ranges, DRHdr); 2890 } else { 2891 uint16_t RegRelFlags = 0; 2892 if (DefRange.IsSubfield) { 2893 RegRelFlags = DefRangeRegisterRelSym::IsSubfieldFlag | 2894 (DefRange.StructOffset 2895 << DefRangeRegisterRelSym::OffsetInParentShift); 2896 } 2897 DefRangeRegisterRelHeader DRHdr; 2898 DRHdr.Register = Reg; 2899 DRHdr.Flags = RegRelFlags; 2900 DRHdr.BasePointerOffset = Offset; 2901 OS.emitCVDefRangeDirective(Ranges, DRHdr); 2902 } 2903 } else { 2904 assert(DefRange.DataOffset == 0 && "unexpected offset into register"); 2905 if (DefRange.IsSubfield) { 2906 DefRangeSubfieldRegisterHeader DRHdr; 2907 DRHdr.Register = DefRange.CVRegister; 2908 DRHdr.MayHaveNoName = 0; 2909 DRHdr.OffsetInParent = DefRange.StructOffset; 2910 OS.emitCVDefRangeDirective(Ranges, DRHdr); 2911 } else { 2912 DefRangeRegisterHeader DRHdr; 2913 DRHdr.Register = DefRange.CVRegister; 2914 DRHdr.MayHaveNoName = 0; 2915 OS.emitCVDefRangeDirective(Ranges, DRHdr); 2916 } 2917 } 2918 } 2919 } 2920 2921 void CodeViewDebug::emitLexicalBlockList(ArrayRef<LexicalBlock *> Blocks, 2922 const FunctionInfo& FI) { 2923 for (LexicalBlock *Block : Blocks) 2924 emitLexicalBlock(*Block, FI); 2925 } 2926 2927 /// Emit an S_BLOCK32 and S_END record pair delimiting the contents of a 2928 /// lexical block scope. 2929 void CodeViewDebug::emitLexicalBlock(const LexicalBlock &Block, 2930 const FunctionInfo& FI) { 2931 MCSymbol *RecordEnd = beginSymbolRecord(SymbolKind::S_BLOCK32); 2932 OS.AddComment("PtrParent"); 2933 OS.emitInt32(0); // PtrParent 2934 OS.AddComment("PtrEnd"); 2935 OS.emitInt32(0); // PtrEnd 2936 OS.AddComment("Code size"); 2937 OS.emitAbsoluteSymbolDiff(Block.End, Block.Begin, 4); // Code Size 2938 OS.AddComment("Function section relative address"); 2939 OS.emitCOFFSecRel32(Block.Begin, /*Offset=*/0); // Func Offset 2940 OS.AddComment("Function section index"); 2941 OS.emitCOFFSectionIndex(FI.Begin); // Func Symbol 2942 OS.AddComment("Lexical block name"); 2943 emitNullTerminatedSymbolName(OS, Block.Name); // Name 2944 endSymbolRecord(RecordEnd); 2945 2946 // Emit variables local to this lexical block. 2947 emitLocalVariableList(FI, Block.Locals); 2948 emitGlobalVariableList(Block.Globals); 2949 2950 // Emit lexical blocks contained within this block. 2951 emitLexicalBlockList(Block.Children, FI); 2952 2953 // Close the lexical block scope. 2954 emitEndSymbolRecord(SymbolKind::S_END); 2955 } 2956 2957 /// Convenience routine for collecting lexical block information for a list 2958 /// of lexical scopes. 2959 void CodeViewDebug::collectLexicalBlockInfo( 2960 SmallVectorImpl<LexicalScope *> &Scopes, 2961 SmallVectorImpl<LexicalBlock *> &Blocks, 2962 SmallVectorImpl<LocalVariable> &Locals, 2963 SmallVectorImpl<CVGlobalVariable> &Globals) { 2964 for (LexicalScope *Scope : Scopes) 2965 collectLexicalBlockInfo(*Scope, Blocks, Locals, Globals); 2966 } 2967 2968 /// Populate the lexical blocks and local variable lists of the parent with 2969 /// information about the specified lexical scope. 2970 void CodeViewDebug::collectLexicalBlockInfo( 2971 LexicalScope &Scope, 2972 SmallVectorImpl<LexicalBlock *> &ParentBlocks, 2973 SmallVectorImpl<LocalVariable> &ParentLocals, 2974 SmallVectorImpl<CVGlobalVariable> &ParentGlobals) { 2975 if (Scope.isAbstractScope()) 2976 return; 2977 2978 // Gather information about the lexical scope including local variables, 2979 // global variables, and address ranges. 2980 bool IgnoreScope = false; 2981 auto LI = ScopeVariables.find(&Scope); 2982 SmallVectorImpl<LocalVariable> *Locals = 2983 LI != ScopeVariables.end() ? &LI->second : nullptr; 2984 auto GI = ScopeGlobals.find(Scope.getScopeNode()); 2985 SmallVectorImpl<CVGlobalVariable> *Globals = 2986 GI != ScopeGlobals.end() ? GI->second.get() : nullptr; 2987 const DILexicalBlock *DILB = dyn_cast<DILexicalBlock>(Scope.getScopeNode()); 2988 const SmallVectorImpl<InsnRange> &Ranges = Scope.getRanges(); 2989 2990 // Ignore lexical scopes which do not contain variables. 2991 if (!Locals && !Globals) 2992 IgnoreScope = true; 2993 2994 // Ignore lexical scopes which are not lexical blocks. 2995 if (!DILB) 2996 IgnoreScope = true; 2997 2998 // Ignore scopes which have too many address ranges to represent in the 2999 // current CodeView format or do not have a valid address range. 3000 // 3001 // For lexical scopes with multiple address ranges you may be tempted to 3002 // construct a single range covering every instruction where the block is 3003 // live and everything in between. Unfortunately, Visual Studio only 3004 // displays variables from the first matching lexical block scope. If the 3005 // first lexical block contains exception handling code or cold code which 3006 // is moved to the bottom of the routine creating a single range covering 3007 // nearly the entire routine, then it will hide all other lexical blocks 3008 // and the variables they contain. 3009 if (Ranges.size() != 1 || !getLabelAfterInsn(Ranges.front().second)) 3010 IgnoreScope = true; 3011 3012 if (IgnoreScope) { 3013 // This scope can be safely ignored and eliminating it will reduce the 3014 // size of the debug information. Be sure to collect any variable and scope 3015 // information from the this scope or any of its children and collapse them 3016 // into the parent scope. 3017 if (Locals) 3018 ParentLocals.append(Locals->begin(), Locals->end()); 3019 if (Globals) 3020 ParentGlobals.append(Globals->begin(), Globals->end()); 3021 collectLexicalBlockInfo(Scope.getChildren(), 3022 ParentBlocks, 3023 ParentLocals, 3024 ParentGlobals); 3025 return; 3026 } 3027 3028 // Create a new CodeView lexical block for this lexical scope. If we've 3029 // seen this DILexicalBlock before then the scope tree is malformed and 3030 // we can handle this gracefully by not processing it a second time. 3031 auto BlockInsertion = CurFn->LexicalBlocks.insert({DILB, LexicalBlock()}); 3032 if (!BlockInsertion.second) 3033 return; 3034 3035 // Create a lexical block containing the variables and collect the the 3036 // lexical block information for the children. 3037 const InsnRange &Range = Ranges.front(); 3038 assert(Range.first && Range.second); 3039 LexicalBlock &Block = BlockInsertion.first->second; 3040 Block.Begin = getLabelBeforeInsn(Range.first); 3041 Block.End = getLabelAfterInsn(Range.second); 3042 assert(Block.Begin && "missing label for scope begin"); 3043 assert(Block.End && "missing label for scope end"); 3044 Block.Name = DILB->getName(); 3045 if (Locals) 3046 Block.Locals = std::move(*Locals); 3047 if (Globals) 3048 Block.Globals = std::move(*Globals); 3049 ParentBlocks.push_back(&Block); 3050 collectLexicalBlockInfo(Scope.getChildren(), 3051 Block.Children, 3052 Block.Locals, 3053 Block.Globals); 3054 } 3055 3056 void CodeViewDebug::endFunctionImpl(const MachineFunction *MF) { 3057 const Function &GV = MF->getFunction(); 3058 assert(FnDebugInfo.count(&GV)); 3059 assert(CurFn == FnDebugInfo[&GV].get()); 3060 3061 collectVariableInfo(GV.getSubprogram()); 3062 3063 // Build the lexical block structure to emit for this routine. 3064 if (LexicalScope *CFS = LScopes.getCurrentFunctionScope()) 3065 collectLexicalBlockInfo(*CFS, 3066 CurFn->ChildBlocks, 3067 CurFn->Locals, 3068 CurFn->Globals); 3069 3070 // Clear the scope and variable information from the map which will not be 3071 // valid after we have finished processing this routine. This also prepares 3072 // the map for the subsequent routine. 3073 ScopeVariables.clear(); 3074 3075 // Don't emit anything if we don't have any line tables. 3076 // Thunks are compiler-generated and probably won't have source correlation. 3077 if (!CurFn->HaveLineInfo && !GV.getSubprogram()->isThunk()) { 3078 FnDebugInfo.erase(&GV); 3079 CurFn = nullptr; 3080 return; 3081 } 3082 3083 // Find heap alloc sites and add to list. 3084 for (const auto &MBB : *MF) { 3085 for (const auto &MI : MBB) { 3086 if (MDNode *MD = MI.getHeapAllocMarker()) { 3087 CurFn->HeapAllocSites.push_back(std::make_tuple(getLabelBeforeInsn(&MI), 3088 getLabelAfterInsn(&MI), 3089 dyn_cast<DIType>(MD))); 3090 } 3091 } 3092 } 3093 3094 bool isThumb = Triple(MMI->getModule()->getTargetTriple()).getArch() == 3095 llvm::Triple::ArchType::thumb; 3096 collectDebugInfoForJumpTables(MF, isThumb); 3097 3098 CurFn->Annotations = MF->getCodeViewAnnotations(); 3099 3100 CurFn->End = Asm->getFunctionEnd(); 3101 3102 CurFn = nullptr; 3103 } 3104 3105 // Usable locations are valid with non-zero line numbers. A line number of zero 3106 // corresponds to optimized code that doesn't have a distinct source location. 3107 // In this case, we try to use the previous or next source location depending on 3108 // the context. 3109 static bool isUsableDebugLoc(DebugLoc DL) { 3110 return DL && DL.getLine() != 0; 3111 } 3112 3113 void CodeViewDebug::beginInstruction(const MachineInstr *MI) { 3114 DebugHandlerBase::beginInstruction(MI); 3115 3116 // Ignore DBG_VALUE and DBG_LABEL locations and function prologue. 3117 if (!Asm || !CurFn || MI->isDebugInstr() || 3118 MI->getFlag(MachineInstr::FrameSetup)) 3119 return; 3120 3121 // If the first instruction of a new MBB has no location, find the first 3122 // instruction with a location and use that. 3123 DebugLoc DL = MI->getDebugLoc(); 3124 if (!isUsableDebugLoc(DL) && MI->getParent() != PrevInstBB) { 3125 for (const auto &NextMI : *MI->getParent()) { 3126 if (NextMI.isDebugInstr()) 3127 continue; 3128 DL = NextMI.getDebugLoc(); 3129 if (isUsableDebugLoc(DL)) 3130 break; 3131 } 3132 // FIXME: Handle the case where the BB has no valid locations. This would 3133 // probably require doing a real dataflow analysis. 3134 } 3135 PrevInstBB = MI->getParent(); 3136 3137 // If we still don't have a debug location, don't record a location. 3138 if (!isUsableDebugLoc(DL)) 3139 return; 3140 3141 maybeRecordLocation(DL, Asm->MF); 3142 } 3143 3144 MCSymbol *CodeViewDebug::beginCVSubsection(DebugSubsectionKind Kind) { 3145 MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(), 3146 *EndLabel = MMI->getContext().createTempSymbol(); 3147 OS.emitInt32(unsigned(Kind)); 3148 OS.AddComment("Subsection size"); 3149 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 4); 3150 OS.emitLabel(BeginLabel); 3151 return EndLabel; 3152 } 3153 3154 void CodeViewDebug::endCVSubsection(MCSymbol *EndLabel) { 3155 OS.emitLabel(EndLabel); 3156 // Every subsection must be aligned to a 4-byte boundary. 3157 OS.emitValueToAlignment(Align(4)); 3158 } 3159 3160 static StringRef getSymbolName(SymbolKind SymKind) { 3161 for (const EnumEntry<SymbolKind> &EE : getSymbolTypeNames()) 3162 if (EE.Value == SymKind) 3163 return EE.Name; 3164 return ""; 3165 } 3166 3167 MCSymbol *CodeViewDebug::beginSymbolRecord(SymbolKind SymKind) { 3168 MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(), 3169 *EndLabel = MMI->getContext().createTempSymbol(); 3170 OS.AddComment("Record length"); 3171 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2); 3172 OS.emitLabel(BeginLabel); 3173 if (OS.isVerboseAsm()) 3174 OS.AddComment("Record kind: " + getSymbolName(SymKind)); 3175 OS.emitInt16(unsigned(SymKind)); 3176 return EndLabel; 3177 } 3178 3179 void CodeViewDebug::endSymbolRecord(MCSymbol *SymEnd) { 3180 // MSVC does not pad out symbol records to four bytes, but LLVM does to avoid 3181 // an extra copy of every symbol record in LLD. This increases object file 3182 // size by less than 1% in the clang build, and is compatible with the Visual 3183 // C++ linker. 3184 OS.emitValueToAlignment(Align(4)); 3185 OS.emitLabel(SymEnd); 3186 } 3187 3188 void CodeViewDebug::emitEndSymbolRecord(SymbolKind EndKind) { 3189 OS.AddComment("Record length"); 3190 OS.emitInt16(2); 3191 if (OS.isVerboseAsm()) 3192 OS.AddComment("Record kind: " + getSymbolName(EndKind)); 3193 OS.emitInt16(uint16_t(EndKind)); // Record Kind 3194 } 3195 3196 void CodeViewDebug::emitDebugInfoForUDTs( 3197 const std::vector<std::pair<std::string, const DIType *>> &UDTs) { 3198 #ifndef NDEBUG 3199 size_t OriginalSize = UDTs.size(); 3200 #endif 3201 for (const auto &UDT : UDTs) { 3202 const DIType *T = UDT.second; 3203 assert(shouldEmitUdt(T)); 3204 MCSymbol *UDTRecordEnd = beginSymbolRecord(SymbolKind::S_UDT); 3205 OS.AddComment("Type"); 3206 OS.emitInt32(getCompleteTypeIndex(T).getIndex()); 3207 assert(OriginalSize == UDTs.size() && 3208 "getCompleteTypeIndex found new UDTs!"); 3209 emitNullTerminatedSymbolName(OS, UDT.first); 3210 endSymbolRecord(UDTRecordEnd); 3211 } 3212 } 3213 3214 void CodeViewDebug::collectGlobalVariableInfo() { 3215 DenseMap<const DIGlobalVariableExpression *, const GlobalVariable *> 3216 GlobalMap; 3217 for (const GlobalVariable &GV : MMI->getModule()->globals()) { 3218 SmallVector<DIGlobalVariableExpression *, 1> GVEs; 3219 GV.getDebugInfo(GVEs); 3220 for (const auto *GVE : GVEs) 3221 GlobalMap[GVE] = &GV; 3222 } 3223 3224 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 3225 for (const MDNode *Node : CUs->operands()) { 3226 const auto *CU = cast<DICompileUnit>(Node); 3227 for (const auto *GVE : CU->getGlobalVariables()) { 3228 const DIGlobalVariable *DIGV = GVE->getVariable(); 3229 const DIExpression *DIE = GVE->getExpression(); 3230 // Don't emit string literals in CodeView, as the only useful parts are 3231 // generally the filename and line number, which isn't possible to output 3232 // in CodeView. String literals should be the only unnamed GlobalVariable 3233 // with debug info. 3234 if (DIGV->getName().empty()) continue; 3235 3236 if ((DIE->getNumElements() == 2) && 3237 (DIE->getElement(0) == dwarf::DW_OP_plus_uconst)) 3238 // Record the constant offset for the variable. 3239 // 3240 // A Fortran common block uses this idiom to encode the offset 3241 // of a variable from the common block's starting address. 3242 CVGlobalVariableOffsets.insert( 3243 std::make_pair(DIGV, DIE->getElement(1))); 3244 3245 // Emit constant global variables in a global symbol section. 3246 if (GlobalMap.count(GVE) == 0 && DIE->isConstant()) { 3247 CVGlobalVariable CVGV = {DIGV, DIE}; 3248 GlobalVariables.emplace_back(std::move(CVGV)); 3249 } 3250 3251 const auto *GV = GlobalMap.lookup(GVE); 3252 if (!GV || GV->isDeclarationForLinker()) 3253 continue; 3254 3255 DIScope *Scope = DIGV->getScope(); 3256 SmallVector<CVGlobalVariable, 1> *VariableList; 3257 if (Scope && isa<DILocalScope>(Scope)) { 3258 // Locate a global variable list for this scope, creating one if 3259 // necessary. 3260 auto Insertion = ScopeGlobals.insert( 3261 {Scope, std::unique_ptr<GlobalVariableList>()}); 3262 if (Insertion.second) 3263 Insertion.first->second = std::make_unique<GlobalVariableList>(); 3264 VariableList = Insertion.first->second.get(); 3265 } else if (GV->hasComdat()) 3266 // Emit this global variable into a COMDAT section. 3267 VariableList = &ComdatVariables; 3268 else 3269 // Emit this global variable in a single global symbol section. 3270 VariableList = &GlobalVariables; 3271 CVGlobalVariable CVGV = {DIGV, GV}; 3272 VariableList->emplace_back(std::move(CVGV)); 3273 } 3274 } 3275 } 3276 3277 void CodeViewDebug::collectDebugInfoForGlobals() { 3278 for (const CVGlobalVariable &CVGV : GlobalVariables) { 3279 const DIGlobalVariable *DIGV = CVGV.DIGV; 3280 const DIScope *Scope = DIGV->getScope(); 3281 getCompleteTypeIndex(DIGV->getType()); 3282 getFullyQualifiedName(Scope, DIGV->getName()); 3283 } 3284 3285 for (const CVGlobalVariable &CVGV : ComdatVariables) { 3286 const DIGlobalVariable *DIGV = CVGV.DIGV; 3287 const DIScope *Scope = DIGV->getScope(); 3288 getCompleteTypeIndex(DIGV->getType()); 3289 getFullyQualifiedName(Scope, DIGV->getName()); 3290 } 3291 } 3292 3293 void CodeViewDebug::emitDebugInfoForGlobals() { 3294 // First, emit all globals that are not in a comdat in a single symbol 3295 // substream. MSVC doesn't like it if the substream is empty, so only open 3296 // it if we have at least one global to emit. 3297 switchToDebugSectionForSymbol(nullptr); 3298 if (!GlobalVariables.empty() || !StaticConstMembers.empty()) { 3299 OS.AddComment("Symbol subsection for globals"); 3300 MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols); 3301 emitGlobalVariableList(GlobalVariables); 3302 emitStaticConstMemberList(); 3303 endCVSubsection(EndLabel); 3304 } 3305 3306 // Second, emit each global that is in a comdat into its own .debug$S 3307 // section along with its own symbol substream. 3308 for (const CVGlobalVariable &CVGV : ComdatVariables) { 3309 const GlobalVariable *GV = cast<const GlobalVariable *>(CVGV.GVInfo); 3310 MCSymbol *GVSym = Asm->getSymbol(GV); 3311 OS.AddComment("Symbol subsection for " + 3312 Twine(GlobalValue::dropLLVMManglingEscape(GV->getName()))); 3313 switchToDebugSectionForSymbol(GVSym); 3314 MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols); 3315 // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. 3316 emitDebugInfoForGlobal(CVGV); 3317 endCVSubsection(EndLabel); 3318 } 3319 } 3320 3321 void CodeViewDebug::emitDebugInfoForRetainedTypes() { 3322 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 3323 for (const MDNode *Node : CUs->operands()) { 3324 for (auto *Ty : cast<DICompileUnit>(Node)->getRetainedTypes()) { 3325 if (DIType *RT = dyn_cast<DIType>(Ty)) { 3326 getTypeIndex(RT); 3327 // FIXME: Add to global/local DTU list. 3328 } 3329 } 3330 } 3331 } 3332 3333 // Emit each global variable in the specified array. 3334 void CodeViewDebug::emitGlobalVariableList(ArrayRef<CVGlobalVariable> Globals) { 3335 for (const CVGlobalVariable &CVGV : Globals) { 3336 // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. 3337 emitDebugInfoForGlobal(CVGV); 3338 } 3339 } 3340 3341 void CodeViewDebug::emitConstantSymbolRecord(const DIType *DTy, APSInt &Value, 3342 const std::string &QualifiedName) { 3343 MCSymbol *SConstantEnd = beginSymbolRecord(SymbolKind::S_CONSTANT); 3344 OS.AddComment("Type"); 3345 OS.emitInt32(getTypeIndex(DTy).getIndex()); 3346 3347 OS.AddComment("Value"); 3348 3349 // Encoded integers shouldn't need more than 10 bytes. 3350 uint8_t Data[10]; 3351 BinaryStreamWriter Writer(Data, llvm::support::endianness::little); 3352 CodeViewRecordIO IO(Writer); 3353 cantFail(IO.mapEncodedInteger(Value)); 3354 StringRef SRef((char *)Data, Writer.getOffset()); 3355 OS.emitBinaryData(SRef); 3356 3357 OS.AddComment("Name"); 3358 emitNullTerminatedSymbolName(OS, QualifiedName); 3359 endSymbolRecord(SConstantEnd); 3360 } 3361 3362 void CodeViewDebug::emitStaticConstMemberList() { 3363 for (const DIDerivedType *DTy : StaticConstMembers) { 3364 const DIScope *Scope = DTy->getScope(); 3365 3366 APSInt Value; 3367 if (const ConstantInt *CI = 3368 dyn_cast_or_null<ConstantInt>(DTy->getConstant())) 3369 Value = APSInt(CI->getValue(), 3370 DebugHandlerBase::isUnsignedDIType(DTy->getBaseType())); 3371 else if (const ConstantFP *CFP = 3372 dyn_cast_or_null<ConstantFP>(DTy->getConstant())) 3373 Value = APSInt(CFP->getValueAPF().bitcastToAPInt(), true); 3374 else 3375 llvm_unreachable("cannot emit a constant without a value"); 3376 3377 emitConstantSymbolRecord(DTy->getBaseType(), Value, 3378 getFullyQualifiedName(Scope, DTy->getName())); 3379 } 3380 } 3381 3382 static bool isFloatDIType(const DIType *Ty) { 3383 if (isa<DICompositeType>(Ty)) 3384 return false; 3385 3386 if (auto *DTy = dyn_cast<DIDerivedType>(Ty)) { 3387 dwarf::Tag T = (dwarf::Tag)Ty->getTag(); 3388 if (T == dwarf::DW_TAG_pointer_type || 3389 T == dwarf::DW_TAG_ptr_to_member_type || 3390 T == dwarf::DW_TAG_reference_type || 3391 T == dwarf::DW_TAG_rvalue_reference_type) 3392 return false; 3393 assert(DTy->getBaseType() && "Expected valid base type"); 3394 return isFloatDIType(DTy->getBaseType()); 3395 } 3396 3397 auto *BTy = cast<DIBasicType>(Ty); 3398 return (BTy->getEncoding() == dwarf::DW_ATE_float); 3399 } 3400 3401 void CodeViewDebug::emitDebugInfoForGlobal(const CVGlobalVariable &CVGV) { 3402 const DIGlobalVariable *DIGV = CVGV.DIGV; 3403 3404 const DIScope *Scope = DIGV->getScope(); 3405 // For static data members, get the scope from the declaration. 3406 if (const auto *MemberDecl = dyn_cast_or_null<DIDerivedType>( 3407 DIGV->getRawStaticDataMemberDeclaration())) 3408 Scope = MemberDecl->getScope(); 3409 // For static local variables and Fortran, the scoping portion is elided 3410 // in its name so that we can reference the variable in the command line 3411 // of the VS debugger. 3412 std::string QualifiedName = 3413 (moduleIsInFortran() || (Scope && isa<DILocalScope>(Scope))) 3414 ? std::string(DIGV->getName()) 3415 : getFullyQualifiedName(Scope, DIGV->getName()); 3416 3417 if (const GlobalVariable *GV = 3418 dyn_cast_if_present<const GlobalVariable *>(CVGV.GVInfo)) { 3419 // DataSym record, see SymbolRecord.h for more info. Thread local data 3420 // happens to have the same format as global data. 3421 MCSymbol *GVSym = Asm->getSymbol(GV); 3422 SymbolKind DataSym = GV->isThreadLocal() 3423 ? (DIGV->isLocalToUnit() ? SymbolKind::S_LTHREAD32 3424 : SymbolKind::S_GTHREAD32) 3425 : (DIGV->isLocalToUnit() ? SymbolKind::S_LDATA32 3426 : SymbolKind::S_GDATA32); 3427 MCSymbol *DataEnd = beginSymbolRecord(DataSym); 3428 OS.AddComment("Type"); 3429 OS.emitInt32(getCompleteTypeIndex(DIGV->getType()).getIndex()); 3430 OS.AddComment("DataOffset"); 3431 3432 uint64_t Offset = 0; 3433 if (CVGlobalVariableOffsets.contains(DIGV)) 3434 // Use the offset seen while collecting info on globals. 3435 Offset = CVGlobalVariableOffsets[DIGV]; 3436 OS.emitCOFFSecRel32(GVSym, Offset); 3437 3438 OS.AddComment("Segment"); 3439 OS.emitCOFFSectionIndex(GVSym); 3440 OS.AddComment("Name"); 3441 const unsigned LengthOfDataRecord = 12; 3442 emitNullTerminatedSymbolName(OS, QualifiedName, LengthOfDataRecord); 3443 endSymbolRecord(DataEnd); 3444 } else { 3445 const DIExpression *DIE = cast<const DIExpression *>(CVGV.GVInfo); 3446 assert(DIE->isConstant() && 3447 "Global constant variables must contain a constant expression."); 3448 3449 // Use unsigned for floats. 3450 bool isUnsigned = isFloatDIType(DIGV->getType()) 3451 ? true 3452 : DebugHandlerBase::isUnsignedDIType(DIGV->getType()); 3453 APSInt Value(APInt(/*BitWidth=*/64, DIE->getElement(1)), isUnsigned); 3454 emitConstantSymbolRecord(DIGV->getType(), Value, QualifiedName); 3455 } 3456 } 3457 3458 void forEachJumpTableBranch( 3459 const MachineFunction *MF, bool isThumb, 3460 const std::function<void(const MachineJumpTableInfo &, const MachineInstr &, 3461 int64_t)> &Callback) { 3462 auto JTI = MF->getJumpTableInfo(); 3463 if (JTI && !JTI->isEmpty()) { 3464 for (const auto &MBB : *MF) { 3465 // Search for indirect branches... 3466 const auto LastMI = MBB.getFirstTerminator(); 3467 if (LastMI != MBB.end() && LastMI->isIndirectBranch()) { 3468 bool foundJTI = false; 3469 if (isThumb) { 3470 // ... that directly use jump table operands. 3471 // NOTE: ARM uses pattern matching to lower its BR_JT SDNode to 3472 // machine instructions, hence inserting a JUMP_TABLE_DEBUG_INFO node 3473 // interferes with this process *but* the resulting pseudo-instruction 3474 // uses a Jump Table operand, so extract the jump table index directly 3475 // from that. 3476 for (const auto &MO : LastMI->operands()) { 3477 if (MO.isJTI()) { 3478 foundJTI = true; 3479 Callback(*JTI, *LastMI, MO.getIndex()); 3480 break; 3481 } 3482 } 3483 } else { 3484 // ... that have jump table debug info. 3485 // NOTE: The debug info is inserted as a JUMP_TABLE_DEBUG_INFO node 3486 // when lowering the BR_JT SDNode to an indirect branch. 3487 for (auto I = MBB.instr_rbegin(), E = MBB.instr_rend(); I != E; ++I) { 3488 if (I->isJumpTableDebugInfo()) { 3489 foundJTI = true; 3490 Callback(*JTI, *LastMI, I->getOperand(0).getImm()); 3491 break; 3492 } 3493 } 3494 } 3495 assert(foundJTI); 3496 } 3497 } 3498 } 3499 } 3500 3501 void CodeViewDebug::discoverJumpTableBranches(const MachineFunction *MF, 3502 bool isThumb) { 3503 forEachJumpTableBranch( 3504 MF, isThumb, 3505 [this](const MachineJumpTableInfo &, const MachineInstr &BranchMI, 3506 int64_t) { requestLabelBeforeInsn(&BranchMI); }); 3507 } 3508 3509 void CodeViewDebug::collectDebugInfoForJumpTables(const MachineFunction *MF, 3510 bool isThumb) { 3511 forEachJumpTableBranch( 3512 MF, isThumb, 3513 [this, MF](const MachineJumpTableInfo &JTI, const MachineInstr &BranchMI, 3514 int64_t JumpTableIndex) { 3515 // For label-difference jump tables, find the base expression. 3516 // Otherwise the jump table uses an absolute address (so no base 3517 // is required). 3518 const MCSymbol *Base; 3519 uint64_t BaseOffset = 0; 3520 const MCSymbol *Branch = getLabelBeforeInsn(&BranchMI); 3521 JumpTableEntrySize EntrySize; 3522 switch (JTI.getEntryKind()) { 3523 case MachineJumpTableInfo::EK_Custom32: 3524 case MachineJumpTableInfo::EK_GPRel32BlockAddress: 3525 case MachineJumpTableInfo::EK_GPRel64BlockAddress: 3526 llvm_unreachable( 3527 "EK_Custom32, EK_GPRel32BlockAddress, and " 3528 "EK_GPRel64BlockAddress should never be emitted for COFF"); 3529 case MachineJumpTableInfo::EK_BlockAddress: 3530 // Each entry is an absolute address. 3531 EntrySize = JumpTableEntrySize::Pointer; 3532 Base = nullptr; 3533 break; 3534 case MachineJumpTableInfo::EK_Inline: 3535 case MachineJumpTableInfo::EK_LabelDifference32: 3536 // Ask the AsmPrinter. 3537 std::tie(Base, BaseOffset, Branch, EntrySize) = 3538 Asm->getCodeViewJumpTableInfo(JumpTableIndex, &BranchMI, Branch); 3539 break; 3540 default: 3541 llvm_unreachable("Unknown JumpTableEntryKind"); 3542 } 3543 3544 CurFn->JumpTables.push_back( 3545 {EntrySize, Base, BaseOffset, Branch, 3546 MF->getJTISymbol(JumpTableIndex, MMI->getContext()), 3547 JTI.getJumpTables()[JumpTableIndex].MBBs.size()}); 3548 }); 3549 } 3550 3551 void CodeViewDebug::emitDebugInfoForJumpTables(const FunctionInfo &FI) { 3552 for (auto JumpTable : FI.JumpTables) { 3553 MCSymbol *JumpTableEnd = beginSymbolRecord(SymbolKind::S_ARMSWITCHTABLE); 3554 if (JumpTable.Base) { 3555 OS.AddComment("Base offset"); 3556 OS.emitCOFFSecRel32(JumpTable.Base, JumpTable.BaseOffset); 3557 OS.AddComment("Base section index"); 3558 OS.emitCOFFSectionIndex(JumpTable.Base); 3559 } else { 3560 OS.AddComment("Base offset"); 3561 OS.emitInt32(0); 3562 OS.AddComment("Base section index"); 3563 OS.emitInt16(0); 3564 } 3565 OS.AddComment("Switch type"); 3566 OS.emitInt16(static_cast<uint16_t>(JumpTable.EntrySize)); 3567 OS.AddComment("Branch offset"); 3568 OS.emitCOFFSecRel32(JumpTable.Branch, /*Offset=*/0); 3569 OS.AddComment("Table offset"); 3570 OS.emitCOFFSecRel32(JumpTable.Table, /*Offset=*/0); 3571 OS.AddComment("Branch section index"); 3572 OS.emitCOFFSectionIndex(JumpTable.Branch); 3573 OS.AddComment("Table section index"); 3574 OS.emitCOFFSectionIndex(JumpTable.Table); 3575 OS.AddComment("Entries count"); 3576 OS.emitInt32(JumpTable.TableSize); 3577 endSymbolRecord(JumpTableEnd); 3578 } 3579 } 3580