1 //===- llvm/lib/CodeGen/AsmPrinter/CodeViewDebug.cpp ----------------------===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // This file contains support for writing Microsoft CodeView debug info. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "CodeViewDebug.h" 14 #include "llvm/ADT/APSInt.h" 15 #include "llvm/ADT/STLExtras.h" 16 #include "llvm/ADT/SmallBitVector.h" 17 #include "llvm/ADT/SmallString.h" 18 #include "llvm/ADT/StringRef.h" 19 #include "llvm/ADT/TinyPtrVector.h" 20 #include "llvm/ADT/Twine.h" 21 #include "llvm/BinaryFormat/COFF.h" 22 #include "llvm/BinaryFormat/Dwarf.h" 23 #include "llvm/CodeGen/AsmPrinter.h" 24 #include "llvm/CodeGen/LexicalScopes.h" 25 #include "llvm/CodeGen/MachineFrameInfo.h" 26 #include "llvm/CodeGen/MachineFunction.h" 27 #include "llvm/CodeGen/MachineInstr.h" 28 #include "llvm/CodeGen/MachineModuleInfo.h" 29 #include "llvm/CodeGen/TargetFrameLowering.h" 30 #include "llvm/CodeGen/TargetLowering.h" 31 #include "llvm/CodeGen/TargetRegisterInfo.h" 32 #include "llvm/CodeGen/TargetSubtargetInfo.h" 33 #include "llvm/Config/llvm-config.h" 34 #include "llvm/DebugInfo/CodeView/CVTypeVisitor.h" 35 #include "llvm/DebugInfo/CodeView/CodeViewRecordIO.h" 36 #include "llvm/DebugInfo/CodeView/ContinuationRecordBuilder.h" 37 #include "llvm/DebugInfo/CodeView/DebugInlineeLinesSubsection.h" 38 #include "llvm/DebugInfo/CodeView/EnumTables.h" 39 #include "llvm/DebugInfo/CodeView/Line.h" 40 #include "llvm/DebugInfo/CodeView/SymbolRecord.h" 41 #include "llvm/DebugInfo/CodeView/TypeRecord.h" 42 #include "llvm/DebugInfo/CodeView/TypeTableCollection.h" 43 #include "llvm/DebugInfo/CodeView/TypeVisitorCallbackPipeline.h" 44 #include "llvm/IR/Constants.h" 45 #include "llvm/IR/DataLayout.h" 46 #include "llvm/IR/DebugInfoMetadata.h" 47 #include "llvm/IR/Function.h" 48 #include "llvm/IR/GlobalValue.h" 49 #include "llvm/IR/GlobalVariable.h" 50 #include "llvm/IR/Metadata.h" 51 #include "llvm/IR/Module.h" 52 #include "llvm/MC/MCAsmInfo.h" 53 #include "llvm/MC/MCContext.h" 54 #include "llvm/MC/MCSectionCOFF.h" 55 #include "llvm/MC/MCStreamer.h" 56 #include "llvm/MC/MCSymbol.h" 57 #include "llvm/Support/BinaryStreamWriter.h" 58 #include "llvm/Support/Casting.h" 59 #include "llvm/Support/Endian.h" 60 #include "llvm/Support/Error.h" 61 #include "llvm/Support/ErrorHandling.h" 62 #include "llvm/Support/FormatVariadic.h" 63 #include "llvm/Support/Path.h" 64 #include "llvm/Support/Program.h" 65 #include "llvm/Support/SMLoc.h" 66 #include "llvm/Support/ScopedPrinter.h" 67 #include "llvm/Target/TargetLoweringObjectFile.h" 68 #include "llvm/Target/TargetMachine.h" 69 #include "llvm/TargetParser/Triple.h" 70 #include <algorithm> 71 #include <cassert> 72 #include <cctype> 73 #include <cstddef> 74 #include <iterator> 75 #include <limits> 76 77 using namespace llvm; 78 using namespace llvm::codeview; 79 80 namespace { 81 class CVMCAdapter : public CodeViewRecordStreamer { 82 public: 83 CVMCAdapter(MCStreamer &OS, TypeCollection &TypeTable) 84 : OS(&OS), TypeTable(TypeTable) {} 85 86 void emitBytes(StringRef Data) override { OS->emitBytes(Data); } 87 88 void emitIntValue(uint64_t Value, unsigned Size) override { 89 OS->emitIntValueInHex(Value, Size); 90 } 91 92 void emitBinaryData(StringRef Data) override { OS->emitBinaryData(Data); } 93 94 void AddComment(const Twine &T) override { OS->AddComment(T); } 95 96 void AddRawComment(const Twine &T) override { OS->emitRawComment(T); } 97 98 bool isVerboseAsm() override { return OS->isVerboseAsm(); } 99 100 std::string getTypeName(TypeIndex TI) override { 101 std::string TypeName; 102 if (!TI.isNoneType()) { 103 if (TI.isSimple()) 104 TypeName = std::string(TypeIndex::simpleTypeName(TI)); 105 else 106 TypeName = std::string(TypeTable.getTypeName(TI)); 107 } 108 return TypeName; 109 } 110 111 private: 112 MCStreamer *OS = nullptr; 113 TypeCollection &TypeTable; 114 }; 115 } // namespace 116 117 static CPUType mapArchToCVCPUType(Triple::ArchType Type) { 118 switch (Type) { 119 case Triple::ArchType::x86: 120 return CPUType::Pentium3; 121 case Triple::ArchType::x86_64: 122 return CPUType::X64; 123 case Triple::ArchType::thumb: 124 // LLVM currently doesn't support Windows CE and so thumb 125 // here is indiscriminately mapped to ARMNT specifically. 126 return CPUType::ARMNT; 127 case Triple::ArchType::aarch64: 128 return CPUType::ARM64; 129 default: 130 report_fatal_error("target architecture doesn't map to a CodeView CPUType"); 131 } 132 } 133 134 CodeViewDebug::CodeViewDebug(AsmPrinter *AP) 135 : DebugHandlerBase(AP), OS(*Asm->OutStreamer), TypeTable(Allocator) {} 136 137 StringRef CodeViewDebug::getFullFilepath(const DIFile *File) { 138 std::string &Filepath = FileToFilepathMap[File]; 139 if (!Filepath.empty()) 140 return Filepath; 141 142 StringRef Dir = File->getDirectory(), Filename = File->getFilename(); 143 144 // If this is a Unix-style path, just use it as is. Don't try to canonicalize 145 // it textually because one of the path components could be a symlink. 146 if (Dir.startswith("/") || Filename.startswith("/")) { 147 if (llvm::sys::path::is_absolute(Filename, llvm::sys::path::Style::posix)) 148 return Filename; 149 Filepath = std::string(Dir); 150 if (Dir.back() != '/') 151 Filepath += '/'; 152 Filepath += Filename; 153 return Filepath; 154 } 155 156 // Clang emits directory and relative filename info into the IR, but CodeView 157 // operates on full paths. We could change Clang to emit full paths too, but 158 // that would increase the IR size and probably not needed for other users. 159 // For now, just concatenate and canonicalize the path here. 160 if (Filename.find(':') == 1) 161 Filepath = std::string(Filename); 162 else 163 Filepath = (Dir + "\\" + Filename).str(); 164 165 // Canonicalize the path. We have to do it textually because we may no longer 166 // have access the file in the filesystem. 167 // First, replace all slashes with backslashes. 168 std::replace(Filepath.begin(), Filepath.end(), '/', '\\'); 169 170 // Remove all "\.\" with "\". 171 size_t Cursor = 0; 172 while ((Cursor = Filepath.find("\\.\\", Cursor)) != std::string::npos) 173 Filepath.erase(Cursor, 2); 174 175 // Replace all "\XXX\..\" with "\". Don't try too hard though as the original 176 // path should be well-formatted, e.g. start with a drive letter, etc. 177 Cursor = 0; 178 while ((Cursor = Filepath.find("\\..\\", Cursor)) != std::string::npos) { 179 // Something's wrong if the path starts with "\..\", abort. 180 if (Cursor == 0) 181 break; 182 183 size_t PrevSlash = Filepath.rfind('\\', Cursor - 1); 184 if (PrevSlash == std::string::npos) 185 // Something's wrong, abort. 186 break; 187 188 Filepath.erase(PrevSlash, Cursor + 3 - PrevSlash); 189 // The next ".." might be following the one we've just erased. 190 Cursor = PrevSlash; 191 } 192 193 // Remove all duplicate backslashes. 194 Cursor = 0; 195 while ((Cursor = Filepath.find("\\\\", Cursor)) != std::string::npos) 196 Filepath.erase(Cursor, 1); 197 198 return Filepath; 199 } 200 201 unsigned CodeViewDebug::maybeRecordFile(const DIFile *F) { 202 StringRef FullPath = getFullFilepath(F); 203 unsigned NextId = FileIdMap.size() + 1; 204 auto Insertion = FileIdMap.insert(std::make_pair(FullPath, NextId)); 205 if (Insertion.second) { 206 // We have to compute the full filepath and emit a .cv_file directive. 207 ArrayRef<uint8_t> ChecksumAsBytes; 208 FileChecksumKind CSKind = FileChecksumKind::None; 209 if (F->getChecksum()) { 210 std::string Checksum = fromHex(F->getChecksum()->Value); 211 void *CKMem = OS.getContext().allocate(Checksum.size(), 1); 212 memcpy(CKMem, Checksum.data(), Checksum.size()); 213 ChecksumAsBytes = ArrayRef<uint8_t>( 214 reinterpret_cast<const uint8_t *>(CKMem), Checksum.size()); 215 switch (F->getChecksum()->Kind) { 216 case DIFile::CSK_MD5: 217 CSKind = FileChecksumKind::MD5; 218 break; 219 case DIFile::CSK_SHA1: 220 CSKind = FileChecksumKind::SHA1; 221 break; 222 case DIFile::CSK_SHA256: 223 CSKind = FileChecksumKind::SHA256; 224 break; 225 } 226 } 227 bool Success = OS.emitCVFileDirective(NextId, FullPath, ChecksumAsBytes, 228 static_cast<unsigned>(CSKind)); 229 (void)Success; 230 assert(Success && ".cv_file directive failed"); 231 } 232 return Insertion.first->second; 233 } 234 235 CodeViewDebug::InlineSite & 236 CodeViewDebug::getInlineSite(const DILocation *InlinedAt, 237 const DISubprogram *Inlinee) { 238 auto SiteInsertion = CurFn->InlineSites.insert({InlinedAt, InlineSite()}); 239 InlineSite *Site = &SiteInsertion.first->second; 240 if (SiteInsertion.second) { 241 unsigned ParentFuncId = CurFn->FuncId; 242 if (const DILocation *OuterIA = InlinedAt->getInlinedAt()) 243 ParentFuncId = 244 getInlineSite(OuterIA, InlinedAt->getScope()->getSubprogram()) 245 .SiteFuncId; 246 247 Site->SiteFuncId = NextFuncId++; 248 OS.emitCVInlineSiteIdDirective( 249 Site->SiteFuncId, ParentFuncId, maybeRecordFile(InlinedAt->getFile()), 250 InlinedAt->getLine(), InlinedAt->getColumn(), SMLoc()); 251 Site->Inlinee = Inlinee; 252 InlinedSubprograms.insert(Inlinee); 253 auto InlineeIdx = getFuncIdForSubprogram(Inlinee); 254 255 if (InlinedAt->getInlinedAt() == nullptr) 256 CurFn->Inlinees.insert(InlineeIdx); 257 } 258 return *Site; 259 } 260 261 static StringRef getPrettyScopeName(const DIScope *Scope) { 262 StringRef ScopeName = Scope->getName(); 263 if (!ScopeName.empty()) 264 return ScopeName; 265 266 switch (Scope->getTag()) { 267 case dwarf::DW_TAG_enumeration_type: 268 case dwarf::DW_TAG_class_type: 269 case dwarf::DW_TAG_structure_type: 270 case dwarf::DW_TAG_union_type: 271 return "<unnamed-tag>"; 272 case dwarf::DW_TAG_namespace: 273 return "`anonymous namespace'"; 274 default: 275 return StringRef(); 276 } 277 } 278 279 const DISubprogram *CodeViewDebug::collectParentScopeNames( 280 const DIScope *Scope, SmallVectorImpl<StringRef> &QualifiedNameComponents) { 281 const DISubprogram *ClosestSubprogram = nullptr; 282 while (Scope != nullptr) { 283 if (ClosestSubprogram == nullptr) 284 ClosestSubprogram = dyn_cast<DISubprogram>(Scope); 285 286 // If a type appears in a scope chain, make sure it gets emitted. The 287 // frontend will be responsible for deciding if this should be a forward 288 // declaration or a complete type. 289 if (const auto *Ty = dyn_cast<DICompositeType>(Scope)) 290 DeferredCompleteTypes.push_back(Ty); 291 292 StringRef ScopeName = getPrettyScopeName(Scope); 293 if (!ScopeName.empty()) 294 QualifiedNameComponents.push_back(ScopeName); 295 Scope = Scope->getScope(); 296 } 297 return ClosestSubprogram; 298 } 299 300 static std::string formatNestedName(ArrayRef<StringRef> QualifiedNameComponents, 301 StringRef TypeName) { 302 std::string FullyQualifiedName; 303 for (StringRef QualifiedNameComponent : 304 llvm::reverse(QualifiedNameComponents)) { 305 FullyQualifiedName.append(std::string(QualifiedNameComponent)); 306 FullyQualifiedName.append("::"); 307 } 308 FullyQualifiedName.append(std::string(TypeName)); 309 return FullyQualifiedName; 310 } 311 312 struct CodeViewDebug::TypeLoweringScope { 313 TypeLoweringScope(CodeViewDebug &CVD) : CVD(CVD) { ++CVD.TypeEmissionLevel; } 314 ~TypeLoweringScope() { 315 // Don't decrement TypeEmissionLevel until after emitting deferred types, so 316 // inner TypeLoweringScopes don't attempt to emit deferred types. 317 if (CVD.TypeEmissionLevel == 1) 318 CVD.emitDeferredCompleteTypes(); 319 --CVD.TypeEmissionLevel; 320 } 321 CodeViewDebug &CVD; 322 }; 323 324 std::string CodeViewDebug::getFullyQualifiedName(const DIScope *Scope, 325 StringRef Name) { 326 // Ensure types in the scope chain are emitted as soon as possible. 327 // This can create otherwise a situation where S_UDTs are emitted while 328 // looping in emitDebugInfoForUDTs. 329 TypeLoweringScope S(*this); 330 SmallVector<StringRef, 5> QualifiedNameComponents; 331 collectParentScopeNames(Scope, QualifiedNameComponents); 332 return formatNestedName(QualifiedNameComponents, Name); 333 } 334 335 std::string CodeViewDebug::getFullyQualifiedName(const DIScope *Ty) { 336 const DIScope *Scope = Ty->getScope(); 337 return getFullyQualifiedName(Scope, getPrettyScopeName(Ty)); 338 } 339 340 TypeIndex CodeViewDebug::getScopeIndex(const DIScope *Scope) { 341 // No scope means global scope and that uses the zero index. 342 // 343 // We also use zero index when the scope is a DISubprogram 344 // to suppress the emission of LF_STRING_ID for the function, 345 // which can trigger a link-time error with the linker in 346 // VS2019 version 16.11.2 or newer. 347 // Note, however, skipping the debug info emission for the DISubprogram 348 // is a temporary fix. The root issue here is that we need to figure out 349 // the proper way to encode a function nested in another function 350 // (as introduced by the Fortran 'contains' keyword) in CodeView. 351 if (!Scope || isa<DIFile>(Scope) || isa<DISubprogram>(Scope)) 352 return TypeIndex(); 353 354 assert(!isa<DIType>(Scope) && "shouldn't make a namespace scope for a type"); 355 356 // Check if we've already translated this scope. 357 auto I = TypeIndices.find({Scope, nullptr}); 358 if (I != TypeIndices.end()) 359 return I->second; 360 361 // Build the fully qualified name of the scope. 362 std::string ScopeName = getFullyQualifiedName(Scope); 363 StringIdRecord SID(TypeIndex(), ScopeName); 364 auto TI = TypeTable.writeLeafType(SID); 365 return recordTypeIndexForDINode(Scope, TI); 366 } 367 368 static StringRef removeTemplateArgs(StringRef Name) { 369 // Remove template args from the display name. Assume that the template args 370 // are the last thing in the name. 371 if (Name.empty() || Name.back() != '>') 372 return Name; 373 374 int OpenBrackets = 0; 375 for (int i = Name.size() - 1; i >= 0; --i) { 376 if (Name[i] == '>') 377 ++OpenBrackets; 378 else if (Name[i] == '<') { 379 --OpenBrackets; 380 if (OpenBrackets == 0) 381 return Name.substr(0, i); 382 } 383 } 384 return Name; 385 } 386 387 TypeIndex CodeViewDebug::getFuncIdForSubprogram(const DISubprogram *SP) { 388 assert(SP); 389 390 // Check if we've already translated this subprogram. 391 auto I = TypeIndices.find({SP, nullptr}); 392 if (I != TypeIndices.end()) 393 return I->second; 394 395 // The display name includes function template arguments. Drop them to match 396 // MSVC. We need to have the template arguments in the DISubprogram name 397 // because they are used in other symbol records, such as S_GPROC32_IDs. 398 StringRef DisplayName = removeTemplateArgs(SP->getName()); 399 400 const DIScope *Scope = SP->getScope(); 401 TypeIndex TI; 402 if (const auto *Class = dyn_cast_or_null<DICompositeType>(Scope)) { 403 // If the scope is a DICompositeType, then this must be a method. Member 404 // function types take some special handling, and require access to the 405 // subprogram. 406 TypeIndex ClassType = getTypeIndex(Class); 407 MemberFuncIdRecord MFuncId(ClassType, getMemberFunctionType(SP, Class), 408 DisplayName); 409 TI = TypeTable.writeLeafType(MFuncId); 410 } else { 411 // Otherwise, this must be a free function. 412 TypeIndex ParentScope = getScopeIndex(Scope); 413 FuncIdRecord FuncId(ParentScope, getTypeIndex(SP->getType()), DisplayName); 414 TI = TypeTable.writeLeafType(FuncId); 415 } 416 417 return recordTypeIndexForDINode(SP, TI); 418 } 419 420 static bool isNonTrivial(const DICompositeType *DCTy) { 421 return ((DCTy->getFlags() & DINode::FlagNonTrivial) == DINode::FlagNonTrivial); 422 } 423 424 static FunctionOptions 425 getFunctionOptions(const DISubroutineType *Ty, 426 const DICompositeType *ClassTy = nullptr, 427 StringRef SPName = StringRef("")) { 428 FunctionOptions FO = FunctionOptions::None; 429 const DIType *ReturnTy = nullptr; 430 if (auto TypeArray = Ty->getTypeArray()) { 431 if (TypeArray.size()) 432 ReturnTy = TypeArray[0]; 433 } 434 435 // Add CxxReturnUdt option to functions that return nontrivial record types 436 // or methods that return record types. 437 if (auto *ReturnDCTy = dyn_cast_or_null<DICompositeType>(ReturnTy)) 438 if (isNonTrivial(ReturnDCTy) || ClassTy) 439 FO |= FunctionOptions::CxxReturnUdt; 440 441 // DISubroutineType is unnamed. Use DISubprogram's i.e. SPName in comparison. 442 if (ClassTy && isNonTrivial(ClassTy) && SPName == ClassTy->getName()) { 443 FO |= FunctionOptions::Constructor; 444 445 // TODO: put the FunctionOptions::ConstructorWithVirtualBases flag. 446 447 } 448 return FO; 449 } 450 451 TypeIndex CodeViewDebug::getMemberFunctionType(const DISubprogram *SP, 452 const DICompositeType *Class) { 453 // Always use the method declaration as the key for the function type. The 454 // method declaration contains the this adjustment. 455 if (SP->getDeclaration()) 456 SP = SP->getDeclaration(); 457 assert(!SP->getDeclaration() && "should use declaration as key"); 458 459 // Key the MemberFunctionRecord into the map as {SP, Class}. It won't collide 460 // with the MemberFuncIdRecord, which is keyed in as {SP, nullptr}. 461 auto I = TypeIndices.find({SP, Class}); 462 if (I != TypeIndices.end()) 463 return I->second; 464 465 // Make sure complete type info for the class is emitted *after* the member 466 // function type, as the complete class type is likely to reference this 467 // member function type. 468 TypeLoweringScope S(*this); 469 const bool IsStaticMethod = (SP->getFlags() & DINode::FlagStaticMember) != 0; 470 471 FunctionOptions FO = getFunctionOptions(SP->getType(), Class, SP->getName()); 472 TypeIndex TI = lowerTypeMemberFunction( 473 SP->getType(), Class, SP->getThisAdjustment(), IsStaticMethod, FO); 474 return recordTypeIndexForDINode(SP, TI, Class); 475 } 476 477 TypeIndex CodeViewDebug::recordTypeIndexForDINode(const DINode *Node, 478 TypeIndex TI, 479 const DIType *ClassTy) { 480 auto InsertResult = TypeIndices.insert({{Node, ClassTy}, TI}); 481 (void)InsertResult; 482 assert(InsertResult.second && "DINode was already assigned a type index"); 483 return TI; 484 } 485 486 unsigned CodeViewDebug::getPointerSizeInBytes() { 487 return MMI->getModule()->getDataLayout().getPointerSizeInBits() / 8; 488 } 489 490 void CodeViewDebug::recordLocalVariable(LocalVariable &&Var, 491 const LexicalScope *LS) { 492 if (const DILocation *InlinedAt = LS->getInlinedAt()) { 493 // This variable was inlined. Associate it with the InlineSite. 494 const DISubprogram *Inlinee = Var.DIVar->getScope()->getSubprogram(); 495 InlineSite &Site = getInlineSite(InlinedAt, Inlinee); 496 Site.InlinedLocals.emplace_back(std::move(Var)); 497 } else { 498 // This variable goes into the corresponding lexical scope. 499 ScopeVariables[LS].emplace_back(std::move(Var)); 500 } 501 } 502 503 static void addLocIfNotPresent(SmallVectorImpl<const DILocation *> &Locs, 504 const DILocation *Loc) { 505 if (!llvm::is_contained(Locs, Loc)) 506 Locs.push_back(Loc); 507 } 508 509 void CodeViewDebug::maybeRecordLocation(const DebugLoc &DL, 510 const MachineFunction *MF) { 511 // Skip this instruction if it has the same location as the previous one. 512 if (!DL || DL == PrevInstLoc) 513 return; 514 515 const DIScope *Scope = DL->getScope(); 516 if (!Scope) 517 return; 518 519 // Skip this line if it is longer than the maximum we can record. 520 LineInfo LI(DL.getLine(), DL.getLine(), /*IsStatement=*/true); 521 if (LI.getStartLine() != DL.getLine() || LI.isAlwaysStepInto() || 522 LI.isNeverStepInto()) 523 return; 524 525 ColumnInfo CI(DL.getCol(), /*EndColumn=*/0); 526 if (CI.getStartColumn() != DL.getCol()) 527 return; 528 529 if (!CurFn->HaveLineInfo) 530 CurFn->HaveLineInfo = true; 531 unsigned FileId = 0; 532 if (PrevInstLoc.get() && PrevInstLoc->getFile() == DL->getFile()) 533 FileId = CurFn->LastFileId; 534 else 535 FileId = CurFn->LastFileId = maybeRecordFile(DL->getFile()); 536 PrevInstLoc = DL; 537 538 unsigned FuncId = CurFn->FuncId; 539 if (const DILocation *SiteLoc = DL->getInlinedAt()) { 540 const DILocation *Loc = DL.get(); 541 542 // If this location was actually inlined from somewhere else, give it the ID 543 // of the inline call site. 544 FuncId = 545 getInlineSite(SiteLoc, Loc->getScope()->getSubprogram()).SiteFuncId; 546 547 // Ensure we have links in the tree of inline call sites. 548 bool FirstLoc = true; 549 while ((SiteLoc = Loc->getInlinedAt())) { 550 InlineSite &Site = 551 getInlineSite(SiteLoc, Loc->getScope()->getSubprogram()); 552 if (!FirstLoc) 553 addLocIfNotPresent(Site.ChildSites, Loc); 554 FirstLoc = false; 555 Loc = SiteLoc; 556 } 557 addLocIfNotPresent(CurFn->ChildSites, Loc); 558 } 559 560 OS.emitCVLocDirective(FuncId, FileId, DL.getLine(), DL.getCol(), 561 /*PrologueEnd=*/false, /*IsStmt=*/false, 562 DL->getFilename(), SMLoc()); 563 } 564 565 void CodeViewDebug::emitCodeViewMagicVersion() { 566 OS.emitValueToAlignment(Align(4)); 567 OS.AddComment("Debug section magic"); 568 OS.emitInt32(COFF::DEBUG_SECTION_MAGIC); 569 } 570 571 static SourceLanguage MapDWLangToCVLang(unsigned DWLang) { 572 switch (DWLang) { 573 case dwarf::DW_LANG_C: 574 case dwarf::DW_LANG_C89: 575 case dwarf::DW_LANG_C99: 576 case dwarf::DW_LANG_C11: 577 return SourceLanguage::C; 578 case dwarf::DW_LANG_C_plus_plus: 579 case dwarf::DW_LANG_C_plus_plus_03: 580 case dwarf::DW_LANG_C_plus_plus_11: 581 case dwarf::DW_LANG_C_plus_plus_14: 582 return SourceLanguage::Cpp; 583 case dwarf::DW_LANG_Fortran77: 584 case dwarf::DW_LANG_Fortran90: 585 case dwarf::DW_LANG_Fortran95: 586 case dwarf::DW_LANG_Fortran03: 587 case dwarf::DW_LANG_Fortran08: 588 return SourceLanguage::Fortran; 589 case dwarf::DW_LANG_Pascal83: 590 return SourceLanguage::Pascal; 591 case dwarf::DW_LANG_Cobol74: 592 case dwarf::DW_LANG_Cobol85: 593 return SourceLanguage::Cobol; 594 case dwarf::DW_LANG_Java: 595 return SourceLanguage::Java; 596 case dwarf::DW_LANG_D: 597 return SourceLanguage::D; 598 case dwarf::DW_LANG_Swift: 599 return SourceLanguage::Swift; 600 case dwarf::DW_LANG_Rust: 601 return SourceLanguage::Rust; 602 case dwarf::DW_LANG_ObjC: 603 return SourceLanguage::ObjC; 604 case dwarf::DW_LANG_ObjC_plus_plus: 605 return SourceLanguage::ObjCpp; 606 default: 607 // There's no CodeView representation for this language, and CV doesn't 608 // have an "unknown" option for the language field, so we'll use MASM, 609 // as it's very low level. 610 return SourceLanguage::Masm; 611 } 612 } 613 614 void CodeViewDebug::beginModule(Module *M) { 615 // If module doesn't have named metadata anchors or COFF debug section 616 // is not available, skip any debug info related stuff. 617 if (!MMI->hasDebugInfo() || 618 !Asm->getObjFileLowering().getCOFFDebugSymbolsSection()) { 619 Asm = nullptr; 620 return; 621 } 622 623 TheCPU = mapArchToCVCPUType(Triple(M->getTargetTriple()).getArch()); 624 625 // Get the current source language. 626 const MDNode *Node = *M->debug_compile_units_begin(); 627 const auto *CU = cast<DICompileUnit>(Node); 628 629 CurrentSourceLanguage = MapDWLangToCVLang(CU->getSourceLanguage()); 630 631 collectGlobalVariableInfo(); 632 633 // Check if we should emit type record hashes. 634 ConstantInt *GH = 635 mdconst::extract_or_null<ConstantInt>(M->getModuleFlag("CodeViewGHash")); 636 EmitDebugGlobalHashes = GH && !GH->isZero(); 637 } 638 639 void CodeViewDebug::endModule() { 640 if (!Asm || !MMI->hasDebugInfo()) 641 return; 642 643 // The COFF .debug$S section consists of several subsections, each starting 644 // with a 4-byte control code (e.g. 0xF1, 0xF2, etc) and then a 4-byte length 645 // of the payload followed by the payload itself. The subsections are 4-byte 646 // aligned. 647 648 // Use the generic .debug$S section, and make a subsection for all the inlined 649 // subprograms. 650 switchToDebugSectionForSymbol(nullptr); 651 652 MCSymbol *CompilerInfo = beginCVSubsection(DebugSubsectionKind::Symbols); 653 emitObjName(); 654 emitCompilerInformation(); 655 endCVSubsection(CompilerInfo); 656 657 emitInlineeLinesSubsection(); 658 659 // Emit per-function debug information. 660 for (auto &P : FnDebugInfo) 661 if (!P.first->isDeclarationForLinker()) 662 emitDebugInfoForFunction(P.first, *P.second); 663 664 // Get types used by globals without emitting anything. 665 // This is meant to collect all static const data members so they can be 666 // emitted as globals. 667 collectDebugInfoForGlobals(); 668 669 // Emit retained types. 670 emitDebugInfoForRetainedTypes(); 671 672 // Emit global variable debug information. 673 setCurrentSubprogram(nullptr); 674 emitDebugInfoForGlobals(); 675 676 // Switch back to the generic .debug$S section after potentially processing 677 // comdat symbol sections. 678 switchToDebugSectionForSymbol(nullptr); 679 680 // Emit UDT records for any types used by global variables. 681 if (!GlobalUDTs.empty()) { 682 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 683 emitDebugInfoForUDTs(GlobalUDTs); 684 endCVSubsection(SymbolsEnd); 685 } 686 687 // This subsection holds a file index to offset in string table table. 688 OS.AddComment("File index to string table offset subsection"); 689 OS.emitCVFileChecksumsDirective(); 690 691 // This subsection holds the string table. 692 OS.AddComment("String table"); 693 OS.emitCVStringTableDirective(); 694 695 // Emit S_BUILDINFO, which points to LF_BUILDINFO. Put this in its own symbol 696 // subsection in the generic .debug$S section at the end. There is no 697 // particular reason for this ordering other than to match MSVC. 698 emitBuildInfo(); 699 700 // Emit type information and hashes last, so that any types we translate while 701 // emitting function info are included. 702 emitTypeInformation(); 703 704 if (EmitDebugGlobalHashes) 705 emitTypeGlobalHashes(); 706 707 clear(); 708 } 709 710 static void 711 emitNullTerminatedSymbolName(MCStreamer &OS, StringRef S, 712 unsigned MaxFixedRecordLength = 0xF00) { 713 // The maximum CV record length is 0xFF00. Most of the strings we emit appear 714 // after a fixed length portion of the record. The fixed length portion should 715 // always be less than 0xF00 (3840) bytes, so truncate the string so that the 716 // overall record size is less than the maximum allowed. 717 SmallString<32> NullTerminatedString( 718 S.take_front(MaxRecordLength - MaxFixedRecordLength - 1)); 719 NullTerminatedString.push_back('\0'); 720 OS.emitBytes(NullTerminatedString); 721 } 722 723 void CodeViewDebug::emitTypeInformation() { 724 if (TypeTable.empty()) 725 return; 726 727 // Start the .debug$T or .debug$P section with 0x4. 728 OS.switchSection(Asm->getObjFileLowering().getCOFFDebugTypesSection()); 729 emitCodeViewMagicVersion(); 730 731 TypeTableCollection Table(TypeTable.records()); 732 TypeVisitorCallbackPipeline Pipeline; 733 734 // To emit type record using Codeview MCStreamer adapter 735 CVMCAdapter CVMCOS(OS, Table); 736 TypeRecordMapping typeMapping(CVMCOS); 737 Pipeline.addCallbackToPipeline(typeMapping); 738 739 std::optional<TypeIndex> B = Table.getFirst(); 740 while (B) { 741 // This will fail if the record data is invalid. 742 CVType Record = Table.getType(*B); 743 744 Error E = codeview::visitTypeRecord(Record, *B, Pipeline); 745 746 if (E) { 747 logAllUnhandledErrors(std::move(E), errs(), "error: "); 748 llvm_unreachable("produced malformed type record"); 749 } 750 751 B = Table.getNext(*B); 752 } 753 } 754 755 void CodeViewDebug::emitTypeGlobalHashes() { 756 if (TypeTable.empty()) 757 return; 758 759 // Start the .debug$H section with the version and hash algorithm, currently 760 // hardcoded to version 0, SHA1. 761 OS.switchSection(Asm->getObjFileLowering().getCOFFGlobalTypeHashesSection()); 762 763 OS.emitValueToAlignment(Align(4)); 764 OS.AddComment("Magic"); 765 OS.emitInt32(COFF::DEBUG_HASHES_SECTION_MAGIC); 766 OS.AddComment("Section Version"); 767 OS.emitInt16(0); 768 OS.AddComment("Hash Algorithm"); 769 OS.emitInt16(uint16_t(GlobalTypeHashAlg::BLAKE3)); 770 771 TypeIndex TI(TypeIndex::FirstNonSimpleIndex); 772 for (const auto &GHR : TypeTable.hashes()) { 773 if (OS.isVerboseAsm()) { 774 // Emit an EOL-comment describing which TypeIndex this hash corresponds 775 // to, as well as the stringified SHA1 hash. 776 SmallString<32> Comment; 777 raw_svector_ostream CommentOS(Comment); 778 CommentOS << formatv("{0:X+} [{1}]", TI.getIndex(), GHR); 779 OS.AddComment(Comment); 780 ++TI; 781 } 782 assert(GHR.Hash.size() == 8); 783 StringRef S(reinterpret_cast<const char *>(GHR.Hash.data()), 784 GHR.Hash.size()); 785 OS.emitBinaryData(S); 786 } 787 } 788 789 void CodeViewDebug::emitObjName() { 790 MCSymbol *CompilerEnd = beginSymbolRecord(SymbolKind::S_OBJNAME); 791 792 StringRef PathRef(Asm->TM.Options.ObjectFilenameForDebug); 793 llvm::SmallString<256> PathStore(PathRef); 794 795 if (PathRef.empty() || PathRef == "-") { 796 // Don't emit the filename if we're writing to stdout or to /dev/null. 797 PathRef = {}; 798 } else { 799 PathRef = PathStore; 800 } 801 802 OS.AddComment("Signature"); 803 OS.emitIntValue(0, 4); 804 805 OS.AddComment("Object name"); 806 emitNullTerminatedSymbolName(OS, PathRef); 807 808 endSymbolRecord(CompilerEnd); 809 } 810 811 namespace { 812 struct Version { 813 int Part[4]; 814 }; 815 } // end anonymous namespace 816 817 // Takes a StringRef like "clang 4.0.0.0 (other nonsense 123)" and parses out 818 // the version number. 819 static Version parseVersion(StringRef Name) { 820 Version V = {{0}}; 821 int N = 0; 822 for (const char C : Name) { 823 if (isdigit(C)) { 824 V.Part[N] *= 10; 825 V.Part[N] += C - '0'; 826 V.Part[N] = 827 std::min<int>(V.Part[N], std::numeric_limits<uint16_t>::max()); 828 } else if (C == '.') { 829 ++N; 830 if (N >= 4) 831 return V; 832 } else if (N > 0) 833 return V; 834 } 835 return V; 836 } 837 838 void CodeViewDebug::emitCompilerInformation() { 839 MCSymbol *CompilerEnd = beginSymbolRecord(SymbolKind::S_COMPILE3); 840 uint32_t Flags = 0; 841 842 // The low byte of the flags indicates the source language. 843 Flags = CurrentSourceLanguage; 844 // TODO: Figure out which other flags need to be set. 845 if (MMI->getModule()->getProfileSummary(/*IsCS*/ false) != nullptr) { 846 Flags |= static_cast<uint32_t>(CompileSym3Flags::PGO); 847 } 848 using ArchType = llvm::Triple::ArchType; 849 ArchType Arch = Triple(MMI->getModule()->getTargetTriple()).getArch(); 850 if (Asm->TM.Options.Hotpatch || Arch == ArchType::thumb || 851 Arch == ArchType::aarch64) { 852 Flags |= static_cast<uint32_t>(CompileSym3Flags::HotPatch); 853 } 854 855 OS.AddComment("Flags and language"); 856 OS.emitInt32(Flags); 857 858 OS.AddComment("CPUType"); 859 OS.emitInt16(static_cast<uint64_t>(TheCPU)); 860 861 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 862 const MDNode *Node = *CUs->operands().begin(); 863 const auto *CU = cast<DICompileUnit>(Node); 864 865 StringRef CompilerVersion = CU->getProducer(); 866 Version FrontVer = parseVersion(CompilerVersion); 867 OS.AddComment("Frontend version"); 868 for (int N : FrontVer.Part) { 869 OS.emitInt16(N); 870 } 871 872 // Some Microsoft tools, like Binscope, expect a backend version number of at 873 // least 8.something, so we'll coerce the LLVM version into a form that 874 // guarantees it'll be big enough without really lying about the version. 875 int Major = 1000 * LLVM_VERSION_MAJOR + 876 10 * LLVM_VERSION_MINOR + 877 LLVM_VERSION_PATCH; 878 // Clamp it for builds that use unusually large version numbers. 879 Major = std::min<int>(Major, std::numeric_limits<uint16_t>::max()); 880 Version BackVer = {{ Major, 0, 0, 0 }}; 881 OS.AddComment("Backend version"); 882 for (int N : BackVer.Part) 883 OS.emitInt16(N); 884 885 OS.AddComment("Null-terminated compiler version string"); 886 emitNullTerminatedSymbolName(OS, CompilerVersion); 887 888 endSymbolRecord(CompilerEnd); 889 } 890 891 static TypeIndex getStringIdTypeIdx(GlobalTypeTableBuilder &TypeTable, 892 StringRef S) { 893 StringIdRecord SIR(TypeIndex(0x0), S); 894 return TypeTable.writeLeafType(SIR); 895 } 896 897 static std::string flattenCommandLine(ArrayRef<std::string> Args, 898 StringRef MainFilename) { 899 std::string FlatCmdLine; 900 raw_string_ostream OS(FlatCmdLine); 901 bool PrintedOneArg = false; 902 if (!StringRef(Args[0]).contains("-cc1")) { 903 llvm::sys::printArg(OS, "-cc1", /*Quote=*/true); 904 PrintedOneArg = true; 905 } 906 for (unsigned i = 0; i < Args.size(); i++) { 907 StringRef Arg = Args[i]; 908 if (Arg.empty()) 909 continue; 910 if (Arg == "-main-file-name" || Arg == "-o") { 911 i++; // Skip this argument and next one. 912 continue; 913 } 914 if (Arg.startswith("-object-file-name") || Arg == MainFilename) 915 continue; 916 // Skip fmessage-length for reproduciability. 917 if (Arg.startswith("-fmessage-length")) 918 continue; 919 if (PrintedOneArg) 920 OS << " "; 921 llvm::sys::printArg(OS, Arg, /*Quote=*/true); 922 PrintedOneArg = true; 923 } 924 OS.flush(); 925 return FlatCmdLine; 926 } 927 928 void CodeViewDebug::emitBuildInfo() { 929 // First, make LF_BUILDINFO. It's a sequence of strings with various bits of 930 // build info. The known prefix is: 931 // - Absolute path of current directory 932 // - Compiler path 933 // - Main source file path, relative to CWD or absolute 934 // - Type server PDB file 935 // - Canonical compiler command line 936 // If frontend and backend compilation are separated (think llc or LTO), it's 937 // not clear if the compiler path should refer to the executable for the 938 // frontend or the backend. Leave it blank for now. 939 TypeIndex BuildInfoArgs[BuildInfoRecord::MaxArgs] = {}; 940 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 941 const MDNode *Node = *CUs->operands().begin(); // FIXME: Multiple CUs. 942 const auto *CU = cast<DICompileUnit>(Node); 943 const DIFile *MainSourceFile = CU->getFile(); 944 BuildInfoArgs[BuildInfoRecord::CurrentDirectory] = 945 getStringIdTypeIdx(TypeTable, MainSourceFile->getDirectory()); 946 BuildInfoArgs[BuildInfoRecord::SourceFile] = 947 getStringIdTypeIdx(TypeTable, MainSourceFile->getFilename()); 948 // FIXME: PDB is intentionally blank unless we implement /Zi type servers. 949 BuildInfoArgs[BuildInfoRecord::TypeServerPDB] = 950 getStringIdTypeIdx(TypeTable, ""); 951 if (Asm->TM.Options.MCOptions.Argv0 != nullptr) { 952 BuildInfoArgs[BuildInfoRecord::BuildTool] = 953 getStringIdTypeIdx(TypeTable, Asm->TM.Options.MCOptions.Argv0); 954 BuildInfoArgs[BuildInfoRecord::CommandLine] = getStringIdTypeIdx( 955 TypeTable, flattenCommandLine(Asm->TM.Options.MCOptions.CommandLineArgs, 956 MainSourceFile->getFilename())); 957 } 958 BuildInfoRecord BIR(BuildInfoArgs); 959 TypeIndex BuildInfoIndex = TypeTable.writeLeafType(BIR); 960 961 // Make a new .debug$S subsection for the S_BUILDINFO record, which points 962 // from the module symbols into the type stream. 963 MCSymbol *BISubsecEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 964 MCSymbol *BIEnd = beginSymbolRecord(SymbolKind::S_BUILDINFO); 965 OS.AddComment("LF_BUILDINFO index"); 966 OS.emitInt32(BuildInfoIndex.getIndex()); 967 endSymbolRecord(BIEnd); 968 endCVSubsection(BISubsecEnd); 969 } 970 971 void CodeViewDebug::emitInlineeLinesSubsection() { 972 if (InlinedSubprograms.empty()) 973 return; 974 975 OS.AddComment("Inlinee lines subsection"); 976 MCSymbol *InlineEnd = beginCVSubsection(DebugSubsectionKind::InlineeLines); 977 978 // We emit the checksum info for files. This is used by debuggers to 979 // determine if a pdb matches the source before loading it. Visual Studio, 980 // for instance, will display a warning that the breakpoints are not valid if 981 // the pdb does not match the source. 982 OS.AddComment("Inlinee lines signature"); 983 OS.emitInt32(unsigned(InlineeLinesSignature::Normal)); 984 985 for (const DISubprogram *SP : InlinedSubprograms) { 986 assert(TypeIndices.count({SP, nullptr})); 987 TypeIndex InlineeIdx = TypeIndices[{SP, nullptr}]; 988 989 OS.addBlankLine(); 990 unsigned FileId = maybeRecordFile(SP->getFile()); 991 OS.AddComment("Inlined function " + SP->getName() + " starts at " + 992 SP->getFilename() + Twine(':') + Twine(SP->getLine())); 993 OS.addBlankLine(); 994 OS.AddComment("Type index of inlined function"); 995 OS.emitInt32(InlineeIdx.getIndex()); 996 OS.AddComment("Offset into filechecksum table"); 997 OS.emitCVFileChecksumOffsetDirective(FileId); 998 OS.AddComment("Starting line number"); 999 OS.emitInt32(SP->getLine()); 1000 } 1001 1002 endCVSubsection(InlineEnd); 1003 } 1004 1005 void CodeViewDebug::emitInlinedCallSite(const FunctionInfo &FI, 1006 const DILocation *InlinedAt, 1007 const InlineSite &Site) { 1008 assert(TypeIndices.count({Site.Inlinee, nullptr})); 1009 TypeIndex InlineeIdx = TypeIndices[{Site.Inlinee, nullptr}]; 1010 1011 // SymbolRecord 1012 MCSymbol *InlineEnd = beginSymbolRecord(SymbolKind::S_INLINESITE); 1013 1014 OS.AddComment("PtrParent"); 1015 OS.emitInt32(0); 1016 OS.AddComment("PtrEnd"); 1017 OS.emitInt32(0); 1018 OS.AddComment("Inlinee type index"); 1019 OS.emitInt32(InlineeIdx.getIndex()); 1020 1021 unsigned FileId = maybeRecordFile(Site.Inlinee->getFile()); 1022 unsigned StartLineNum = Site.Inlinee->getLine(); 1023 1024 OS.emitCVInlineLinetableDirective(Site.SiteFuncId, FileId, StartLineNum, 1025 FI.Begin, FI.End); 1026 1027 endSymbolRecord(InlineEnd); 1028 1029 emitLocalVariableList(FI, Site.InlinedLocals); 1030 1031 // Recurse on child inlined call sites before closing the scope. 1032 for (const DILocation *ChildSite : Site.ChildSites) { 1033 auto I = FI.InlineSites.find(ChildSite); 1034 assert(I != FI.InlineSites.end() && 1035 "child site not in function inline site map"); 1036 emitInlinedCallSite(FI, ChildSite, I->second); 1037 } 1038 1039 // Close the scope. 1040 emitEndSymbolRecord(SymbolKind::S_INLINESITE_END); 1041 } 1042 1043 void CodeViewDebug::switchToDebugSectionForSymbol(const MCSymbol *GVSym) { 1044 // If we have a symbol, it may be in a section that is COMDAT. If so, find the 1045 // comdat key. A section may be comdat because of -ffunction-sections or 1046 // because it is comdat in the IR. 1047 MCSectionCOFF *GVSec = 1048 GVSym ? dyn_cast<MCSectionCOFF>(&GVSym->getSection()) : nullptr; 1049 const MCSymbol *KeySym = GVSec ? GVSec->getCOMDATSymbol() : nullptr; 1050 1051 MCSectionCOFF *DebugSec = cast<MCSectionCOFF>( 1052 Asm->getObjFileLowering().getCOFFDebugSymbolsSection()); 1053 DebugSec = OS.getContext().getAssociativeCOFFSection(DebugSec, KeySym); 1054 1055 OS.switchSection(DebugSec); 1056 1057 // Emit the magic version number if this is the first time we've switched to 1058 // this section. 1059 if (ComdatDebugSections.insert(DebugSec).second) 1060 emitCodeViewMagicVersion(); 1061 } 1062 1063 // Emit an S_THUNK32/S_END symbol pair for a thunk routine. 1064 // The only supported thunk ordinal is currently the standard type. 1065 void CodeViewDebug::emitDebugInfoForThunk(const Function *GV, 1066 FunctionInfo &FI, 1067 const MCSymbol *Fn) { 1068 std::string FuncName = 1069 std::string(GlobalValue::dropLLVMManglingEscape(GV->getName())); 1070 const ThunkOrdinal ordinal = ThunkOrdinal::Standard; // Only supported kind. 1071 1072 OS.AddComment("Symbol subsection for " + Twine(FuncName)); 1073 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 1074 1075 // Emit S_THUNK32 1076 MCSymbol *ThunkRecordEnd = beginSymbolRecord(SymbolKind::S_THUNK32); 1077 OS.AddComment("PtrParent"); 1078 OS.emitInt32(0); 1079 OS.AddComment("PtrEnd"); 1080 OS.emitInt32(0); 1081 OS.AddComment("PtrNext"); 1082 OS.emitInt32(0); 1083 OS.AddComment("Thunk section relative address"); 1084 OS.emitCOFFSecRel32(Fn, /*Offset=*/0); 1085 OS.AddComment("Thunk section index"); 1086 OS.emitCOFFSectionIndex(Fn); 1087 OS.AddComment("Code size"); 1088 OS.emitAbsoluteSymbolDiff(FI.End, Fn, 2); 1089 OS.AddComment("Ordinal"); 1090 OS.emitInt8(unsigned(ordinal)); 1091 OS.AddComment("Function name"); 1092 emitNullTerminatedSymbolName(OS, FuncName); 1093 // Additional fields specific to the thunk ordinal would go here. 1094 endSymbolRecord(ThunkRecordEnd); 1095 1096 // Local variables/inlined routines are purposely omitted here. The point of 1097 // marking this as a thunk is so Visual Studio will NOT stop in this routine. 1098 1099 // Emit S_PROC_ID_END 1100 emitEndSymbolRecord(SymbolKind::S_PROC_ID_END); 1101 1102 endCVSubsection(SymbolsEnd); 1103 } 1104 1105 void CodeViewDebug::emitDebugInfoForFunction(const Function *GV, 1106 FunctionInfo &FI) { 1107 // For each function there is a separate subsection which holds the PC to 1108 // file:line table. 1109 const MCSymbol *Fn = Asm->getSymbol(GV); 1110 assert(Fn); 1111 1112 // Switch to the to a comdat section, if appropriate. 1113 switchToDebugSectionForSymbol(Fn); 1114 1115 std::string FuncName; 1116 auto *SP = GV->getSubprogram(); 1117 assert(SP); 1118 setCurrentSubprogram(SP); 1119 1120 if (SP->isThunk()) { 1121 emitDebugInfoForThunk(GV, FI, Fn); 1122 return; 1123 } 1124 1125 // If we have a display name, build the fully qualified name by walking the 1126 // chain of scopes. 1127 if (!SP->getName().empty()) 1128 FuncName = getFullyQualifiedName(SP->getScope(), SP->getName()); 1129 1130 // If our DISubprogram name is empty, use the mangled name. 1131 if (FuncName.empty()) 1132 FuncName = std::string(GlobalValue::dropLLVMManglingEscape(GV->getName())); 1133 1134 // Emit FPO data, but only on 32-bit x86. No other platforms use it. 1135 if (Triple(MMI->getModule()->getTargetTriple()).getArch() == Triple::x86) 1136 OS.emitCVFPOData(Fn); 1137 1138 // Emit a symbol subsection, required by VS2012+ to find function boundaries. 1139 OS.AddComment("Symbol subsection for " + Twine(FuncName)); 1140 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 1141 { 1142 SymbolKind ProcKind = GV->hasLocalLinkage() ? SymbolKind::S_LPROC32_ID 1143 : SymbolKind::S_GPROC32_ID; 1144 MCSymbol *ProcRecordEnd = beginSymbolRecord(ProcKind); 1145 1146 // These fields are filled in by tools like CVPACK which run after the fact. 1147 OS.AddComment("PtrParent"); 1148 OS.emitInt32(0); 1149 OS.AddComment("PtrEnd"); 1150 OS.emitInt32(0); 1151 OS.AddComment("PtrNext"); 1152 OS.emitInt32(0); 1153 // This is the important bit that tells the debugger where the function 1154 // code is located and what's its size: 1155 OS.AddComment("Code size"); 1156 OS.emitAbsoluteSymbolDiff(FI.End, Fn, 4); 1157 OS.AddComment("Offset after prologue"); 1158 OS.emitInt32(0); 1159 OS.AddComment("Offset before epilogue"); 1160 OS.emitInt32(0); 1161 OS.AddComment("Function type index"); 1162 OS.emitInt32(getFuncIdForSubprogram(GV->getSubprogram()).getIndex()); 1163 OS.AddComment("Function section relative address"); 1164 OS.emitCOFFSecRel32(Fn, /*Offset=*/0); 1165 OS.AddComment("Function section index"); 1166 OS.emitCOFFSectionIndex(Fn); 1167 OS.AddComment("Flags"); 1168 ProcSymFlags ProcFlags = ProcSymFlags::HasOptimizedDebugInfo; 1169 if (FI.HasFramePointer) 1170 ProcFlags |= ProcSymFlags::HasFP; 1171 if (GV->hasFnAttribute(Attribute::NoReturn)) 1172 ProcFlags |= ProcSymFlags::IsNoReturn; 1173 if (GV->hasFnAttribute(Attribute::NoInline)) 1174 ProcFlags |= ProcSymFlags::IsNoInline; 1175 OS.emitInt8(static_cast<uint8_t>(ProcFlags)); 1176 // Emit the function display name as a null-terminated string. 1177 OS.AddComment("Function name"); 1178 // Truncate the name so we won't overflow the record length field. 1179 emitNullTerminatedSymbolName(OS, FuncName); 1180 endSymbolRecord(ProcRecordEnd); 1181 1182 MCSymbol *FrameProcEnd = beginSymbolRecord(SymbolKind::S_FRAMEPROC); 1183 // Subtract out the CSR size since MSVC excludes that and we include it. 1184 OS.AddComment("FrameSize"); 1185 OS.emitInt32(FI.FrameSize - FI.CSRSize); 1186 OS.AddComment("Padding"); 1187 OS.emitInt32(0); 1188 OS.AddComment("Offset of padding"); 1189 OS.emitInt32(0); 1190 OS.AddComment("Bytes of callee saved registers"); 1191 OS.emitInt32(FI.CSRSize); 1192 OS.AddComment("Exception handler offset"); 1193 OS.emitInt32(0); 1194 OS.AddComment("Exception handler section"); 1195 OS.emitInt16(0); 1196 OS.AddComment("Flags (defines frame register)"); 1197 OS.emitInt32(uint32_t(FI.FrameProcOpts)); 1198 endSymbolRecord(FrameProcEnd); 1199 1200 emitInlinees(FI.Inlinees); 1201 emitLocalVariableList(FI, FI.Locals); 1202 emitGlobalVariableList(FI.Globals); 1203 emitLexicalBlockList(FI.ChildBlocks, FI); 1204 1205 // Emit inlined call site information. Only emit functions inlined directly 1206 // into the parent function. We'll emit the other sites recursively as part 1207 // of their parent inline site. 1208 for (const DILocation *InlinedAt : FI.ChildSites) { 1209 auto I = FI.InlineSites.find(InlinedAt); 1210 assert(I != FI.InlineSites.end() && 1211 "child site not in function inline site map"); 1212 emitInlinedCallSite(FI, InlinedAt, I->second); 1213 } 1214 1215 for (auto Annot : FI.Annotations) { 1216 MCSymbol *Label = Annot.first; 1217 MDTuple *Strs = cast<MDTuple>(Annot.second); 1218 MCSymbol *AnnotEnd = beginSymbolRecord(SymbolKind::S_ANNOTATION); 1219 OS.emitCOFFSecRel32(Label, /*Offset=*/0); 1220 // FIXME: Make sure we don't overflow the max record size. 1221 OS.emitCOFFSectionIndex(Label); 1222 OS.emitInt16(Strs->getNumOperands()); 1223 for (Metadata *MD : Strs->operands()) { 1224 // MDStrings are null terminated, so we can do EmitBytes and get the 1225 // nice .asciz directive. 1226 StringRef Str = cast<MDString>(MD)->getString(); 1227 assert(Str.data()[Str.size()] == '\0' && "non-nullterminated MDString"); 1228 OS.emitBytes(StringRef(Str.data(), Str.size() + 1)); 1229 } 1230 endSymbolRecord(AnnotEnd); 1231 } 1232 1233 for (auto HeapAllocSite : FI.HeapAllocSites) { 1234 const MCSymbol *BeginLabel = std::get<0>(HeapAllocSite); 1235 const MCSymbol *EndLabel = std::get<1>(HeapAllocSite); 1236 const DIType *DITy = std::get<2>(HeapAllocSite); 1237 MCSymbol *HeapAllocEnd = beginSymbolRecord(SymbolKind::S_HEAPALLOCSITE); 1238 OS.AddComment("Call site offset"); 1239 OS.emitCOFFSecRel32(BeginLabel, /*Offset=*/0); 1240 OS.AddComment("Call site section index"); 1241 OS.emitCOFFSectionIndex(BeginLabel); 1242 OS.AddComment("Call instruction length"); 1243 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2); 1244 OS.AddComment("Type index"); 1245 OS.emitInt32(getCompleteTypeIndex(DITy).getIndex()); 1246 endSymbolRecord(HeapAllocEnd); 1247 } 1248 1249 if (SP != nullptr) 1250 emitDebugInfoForUDTs(LocalUDTs); 1251 1252 emitDebugInfoForJumpTables(FI); 1253 1254 // We're done with this function. 1255 emitEndSymbolRecord(SymbolKind::S_PROC_ID_END); 1256 } 1257 endCVSubsection(SymbolsEnd); 1258 1259 // We have an assembler directive that takes care of the whole line table. 1260 OS.emitCVLinetableDirective(FI.FuncId, Fn, FI.End); 1261 } 1262 1263 CodeViewDebug::LocalVarDef 1264 CodeViewDebug::createDefRangeMem(uint16_t CVRegister, int Offset) { 1265 LocalVarDef DR; 1266 DR.InMemory = -1; 1267 DR.DataOffset = Offset; 1268 assert(DR.DataOffset == Offset && "truncation"); 1269 DR.IsSubfield = 0; 1270 DR.StructOffset = 0; 1271 DR.CVRegister = CVRegister; 1272 return DR; 1273 } 1274 1275 void CodeViewDebug::collectVariableInfoFromMFTable( 1276 DenseSet<InlinedEntity> &Processed) { 1277 const MachineFunction &MF = *Asm->MF; 1278 const TargetSubtargetInfo &TSI = MF.getSubtarget(); 1279 const TargetFrameLowering *TFI = TSI.getFrameLowering(); 1280 const TargetRegisterInfo *TRI = TSI.getRegisterInfo(); 1281 1282 for (const MachineFunction::VariableDbgInfo &VI : 1283 MF.getInStackSlotVariableDbgInfo()) { 1284 if (!VI.Var) 1285 continue; 1286 assert(VI.Var->isValidLocationForIntrinsic(VI.Loc) && 1287 "Expected inlined-at fields to agree"); 1288 1289 Processed.insert(InlinedEntity(VI.Var, VI.Loc->getInlinedAt())); 1290 LexicalScope *Scope = LScopes.findLexicalScope(VI.Loc); 1291 1292 // If variable scope is not found then skip this variable. 1293 if (!Scope) 1294 continue; 1295 1296 // If the variable has an attached offset expression, extract it. 1297 // FIXME: Try to handle DW_OP_deref as well. 1298 int64_t ExprOffset = 0; 1299 bool Deref = false; 1300 if (VI.Expr) { 1301 // If there is one DW_OP_deref element, use offset of 0 and keep going. 1302 if (VI.Expr->getNumElements() == 1 && 1303 VI.Expr->getElement(0) == llvm::dwarf::DW_OP_deref) 1304 Deref = true; 1305 else if (!VI.Expr->extractIfOffset(ExprOffset)) 1306 continue; 1307 } 1308 1309 // Get the frame register used and the offset. 1310 Register FrameReg; 1311 StackOffset FrameOffset = 1312 TFI->getFrameIndexReference(*Asm->MF, VI.getStackSlot(), FrameReg); 1313 uint16_t CVReg = TRI->getCodeViewRegNum(FrameReg); 1314 1315 assert(!FrameOffset.getScalable() && 1316 "Frame offsets with a scalable component are not supported"); 1317 1318 // Calculate the label ranges. 1319 LocalVarDef DefRange = 1320 createDefRangeMem(CVReg, FrameOffset.getFixed() + ExprOffset); 1321 1322 LocalVariable Var; 1323 Var.DIVar = VI.Var; 1324 1325 for (const InsnRange &Range : Scope->getRanges()) { 1326 const MCSymbol *Begin = getLabelBeforeInsn(Range.first); 1327 const MCSymbol *End = getLabelAfterInsn(Range.second); 1328 End = End ? End : Asm->getFunctionEnd(); 1329 Var.DefRanges[DefRange].emplace_back(Begin, End); 1330 } 1331 1332 if (Deref) 1333 Var.UseReferenceType = true; 1334 1335 recordLocalVariable(std::move(Var), Scope); 1336 } 1337 } 1338 1339 static bool canUseReferenceType(const DbgVariableLocation &Loc) { 1340 return !Loc.LoadChain.empty() && Loc.LoadChain.back() == 0; 1341 } 1342 1343 static bool needsReferenceType(const DbgVariableLocation &Loc) { 1344 return Loc.LoadChain.size() == 2 && Loc.LoadChain.back() == 0; 1345 } 1346 1347 void CodeViewDebug::calculateRanges( 1348 LocalVariable &Var, const DbgValueHistoryMap::Entries &Entries) { 1349 const TargetRegisterInfo *TRI = Asm->MF->getSubtarget().getRegisterInfo(); 1350 1351 // Calculate the definition ranges. 1352 for (auto I = Entries.begin(), E = Entries.end(); I != E; ++I) { 1353 const auto &Entry = *I; 1354 if (!Entry.isDbgValue()) 1355 continue; 1356 const MachineInstr *DVInst = Entry.getInstr(); 1357 assert(DVInst->isDebugValue() && "Invalid History entry"); 1358 // FIXME: Find a way to represent constant variables, since they are 1359 // relatively common. 1360 std::optional<DbgVariableLocation> Location = 1361 DbgVariableLocation::extractFromMachineInstruction(*DVInst); 1362 if (!Location) 1363 { 1364 // When we don't have a location this is usually because LLVM has 1365 // transformed it into a constant and we only have an llvm.dbg.value. We 1366 // can't represent these well in CodeView since S_LOCAL only works on 1367 // registers and memory locations. Instead, we will pretend this to be a 1368 // constant value to at least have it show up in the debugger. 1369 auto Op = DVInst->getDebugOperand(0); 1370 if (Op.isImm()) 1371 Var.ConstantValue = APSInt(APInt(64, Op.getImm()), false); 1372 continue; 1373 } 1374 1375 // CodeView can only express variables in register and variables in memory 1376 // at a constant offset from a register. However, for variables passed 1377 // indirectly by pointer, it is common for that pointer to be spilled to a 1378 // stack location. For the special case of one offseted load followed by a 1379 // zero offset load (a pointer spilled to the stack), we change the type of 1380 // the local variable from a value type to a reference type. This tricks the 1381 // debugger into doing the load for us. 1382 if (Var.UseReferenceType) { 1383 // We're using a reference type. Drop the last zero offset load. 1384 if (canUseReferenceType(*Location)) 1385 Location->LoadChain.pop_back(); 1386 else 1387 continue; 1388 } else if (needsReferenceType(*Location)) { 1389 // This location can't be expressed without switching to a reference type. 1390 // Start over using that. 1391 Var.UseReferenceType = true; 1392 Var.DefRanges.clear(); 1393 calculateRanges(Var, Entries); 1394 return; 1395 } 1396 1397 // We can only handle a register or an offseted load of a register. 1398 if (Location->Register == 0 || Location->LoadChain.size() > 1) 1399 continue; 1400 1401 LocalVarDef DR; 1402 DR.CVRegister = TRI->getCodeViewRegNum(Location->Register); 1403 DR.InMemory = !Location->LoadChain.empty(); 1404 DR.DataOffset = 1405 !Location->LoadChain.empty() ? Location->LoadChain.back() : 0; 1406 if (Location->FragmentInfo) { 1407 DR.IsSubfield = true; 1408 DR.StructOffset = Location->FragmentInfo->OffsetInBits / 8; 1409 } else { 1410 DR.IsSubfield = false; 1411 DR.StructOffset = 0; 1412 } 1413 1414 // Compute the label range. 1415 const MCSymbol *Begin = getLabelBeforeInsn(Entry.getInstr()); 1416 const MCSymbol *End; 1417 if (Entry.getEndIndex() != DbgValueHistoryMap::NoEntry) { 1418 auto &EndingEntry = Entries[Entry.getEndIndex()]; 1419 End = EndingEntry.isDbgValue() 1420 ? getLabelBeforeInsn(EndingEntry.getInstr()) 1421 : getLabelAfterInsn(EndingEntry.getInstr()); 1422 } else 1423 End = Asm->getFunctionEnd(); 1424 1425 // If the last range end is our begin, just extend the last range. 1426 // Otherwise make a new range. 1427 SmallVectorImpl<std::pair<const MCSymbol *, const MCSymbol *>> &R = 1428 Var.DefRanges[DR]; 1429 if (!R.empty() && R.back().second == Begin) 1430 R.back().second = End; 1431 else 1432 R.emplace_back(Begin, End); 1433 1434 // FIXME: Do more range combining. 1435 } 1436 } 1437 1438 void CodeViewDebug::collectVariableInfo(const DISubprogram *SP) { 1439 DenseSet<InlinedEntity> Processed; 1440 // Grab the variable info that was squirreled away in the MMI side-table. 1441 collectVariableInfoFromMFTable(Processed); 1442 1443 for (const auto &I : DbgValues) { 1444 InlinedEntity IV = I.first; 1445 if (Processed.count(IV)) 1446 continue; 1447 const DILocalVariable *DIVar = cast<DILocalVariable>(IV.first); 1448 const DILocation *InlinedAt = IV.second; 1449 1450 // Instruction ranges, specifying where IV is accessible. 1451 const auto &Entries = I.second; 1452 1453 LexicalScope *Scope = nullptr; 1454 if (InlinedAt) 1455 Scope = LScopes.findInlinedScope(DIVar->getScope(), InlinedAt); 1456 else 1457 Scope = LScopes.findLexicalScope(DIVar->getScope()); 1458 // If variable scope is not found then skip this variable. 1459 if (!Scope) 1460 continue; 1461 1462 LocalVariable Var; 1463 Var.DIVar = DIVar; 1464 1465 calculateRanges(Var, Entries); 1466 recordLocalVariable(std::move(Var), Scope); 1467 } 1468 } 1469 1470 void CodeViewDebug::beginFunctionImpl(const MachineFunction *MF) { 1471 const TargetSubtargetInfo &TSI = MF->getSubtarget(); 1472 const TargetRegisterInfo *TRI = TSI.getRegisterInfo(); 1473 const MachineFrameInfo &MFI = MF->getFrameInfo(); 1474 const Function &GV = MF->getFunction(); 1475 auto Insertion = FnDebugInfo.insert({&GV, std::make_unique<FunctionInfo>()}); 1476 assert(Insertion.second && "function already has info"); 1477 CurFn = Insertion.first->second.get(); 1478 CurFn->FuncId = NextFuncId++; 1479 CurFn->Begin = Asm->getFunctionBegin(); 1480 1481 // The S_FRAMEPROC record reports the stack size, and how many bytes of 1482 // callee-saved registers were used. For targets that don't use a PUSH 1483 // instruction (AArch64), this will be zero. 1484 CurFn->CSRSize = MFI.getCVBytesOfCalleeSavedRegisters(); 1485 CurFn->FrameSize = MFI.getStackSize(); 1486 CurFn->OffsetAdjustment = MFI.getOffsetAdjustment(); 1487 CurFn->HasStackRealignment = TRI->hasStackRealignment(*MF); 1488 1489 // For this function S_FRAMEPROC record, figure out which codeview register 1490 // will be the frame pointer. 1491 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::None; // None. 1492 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::None; // None. 1493 if (CurFn->FrameSize > 0) { 1494 if (!TSI.getFrameLowering()->hasFP(*MF)) { 1495 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr; 1496 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::StackPtr; 1497 } else { 1498 CurFn->HasFramePointer = true; 1499 // If there is an FP, parameters are always relative to it. 1500 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::FramePtr; 1501 if (CurFn->HasStackRealignment) { 1502 // If the stack needs realignment, locals are relative to SP or VFRAME. 1503 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr; 1504 } else { 1505 // Otherwise, locals are relative to EBP, and we probably have VLAs or 1506 // other stack adjustments. 1507 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::FramePtr; 1508 } 1509 } 1510 } 1511 1512 // Compute other frame procedure options. 1513 FrameProcedureOptions FPO = FrameProcedureOptions::None; 1514 if (MFI.hasVarSizedObjects()) 1515 FPO |= FrameProcedureOptions::HasAlloca; 1516 if (MF->exposesReturnsTwice()) 1517 FPO |= FrameProcedureOptions::HasSetJmp; 1518 // FIXME: Set HasLongJmp if we ever track that info. 1519 if (MF->hasInlineAsm()) 1520 FPO |= FrameProcedureOptions::HasInlineAssembly; 1521 if (GV.hasPersonalityFn()) { 1522 if (isAsynchronousEHPersonality( 1523 classifyEHPersonality(GV.getPersonalityFn()))) 1524 FPO |= FrameProcedureOptions::HasStructuredExceptionHandling; 1525 else 1526 FPO |= FrameProcedureOptions::HasExceptionHandling; 1527 } 1528 if (GV.hasFnAttribute(Attribute::InlineHint)) 1529 FPO |= FrameProcedureOptions::MarkedInline; 1530 if (GV.hasFnAttribute(Attribute::Naked)) 1531 FPO |= FrameProcedureOptions::Naked; 1532 if (MFI.hasStackProtectorIndex()) { 1533 FPO |= FrameProcedureOptions::SecurityChecks; 1534 if (GV.hasFnAttribute(Attribute::StackProtectStrong) || 1535 GV.hasFnAttribute(Attribute::StackProtectReq)) { 1536 FPO |= FrameProcedureOptions::StrictSecurityChecks; 1537 } 1538 } else if (!GV.hasStackProtectorFnAttr()) { 1539 // __declspec(safebuffers) disables stack guards. 1540 FPO |= FrameProcedureOptions::SafeBuffers; 1541 } 1542 FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedLocalFramePtrReg) << 14U); 1543 FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedParamFramePtrReg) << 16U); 1544 if (Asm->TM.getOptLevel() != CodeGenOptLevel::None && !GV.hasOptSize() && 1545 !GV.hasOptNone()) 1546 FPO |= FrameProcedureOptions::OptimizedForSpeed; 1547 if (GV.hasProfileData()) { 1548 FPO |= FrameProcedureOptions::ValidProfileCounts; 1549 FPO |= FrameProcedureOptions::ProfileGuidedOptimization; 1550 } 1551 // FIXME: Set GuardCfg when it is implemented. 1552 CurFn->FrameProcOpts = FPO; 1553 1554 OS.emitCVFuncIdDirective(CurFn->FuncId); 1555 1556 // Find the end of the function prolog. First known non-DBG_VALUE and 1557 // non-frame setup location marks the beginning of the function body. 1558 // FIXME: is there a simpler a way to do this? Can we just search 1559 // for the first instruction of the function, not the last of the prolog? 1560 DebugLoc PrologEndLoc; 1561 bool EmptyPrologue = true; 1562 for (const auto &MBB : *MF) { 1563 for (const auto &MI : MBB) { 1564 if (!MI.isMetaInstruction() && !MI.getFlag(MachineInstr::FrameSetup) && 1565 MI.getDebugLoc()) { 1566 PrologEndLoc = MI.getDebugLoc(); 1567 break; 1568 } else if (!MI.isMetaInstruction()) { 1569 EmptyPrologue = false; 1570 } 1571 } 1572 } 1573 1574 // Record beginning of function if we have a non-empty prologue. 1575 if (PrologEndLoc && !EmptyPrologue) { 1576 DebugLoc FnStartDL = PrologEndLoc.getFnDebugLoc(); 1577 maybeRecordLocation(FnStartDL, MF); 1578 } 1579 1580 // Find heap alloc sites and emit labels around them. 1581 for (const auto &MBB : *MF) { 1582 for (const auto &MI : MBB) { 1583 if (MI.getHeapAllocMarker()) { 1584 requestLabelBeforeInsn(&MI); 1585 requestLabelAfterInsn(&MI); 1586 } 1587 } 1588 } 1589 1590 // Mark branches that may potentially be using jump tables with labels. 1591 bool isThumb = Triple(MMI->getModule()->getTargetTriple()).getArch() == 1592 llvm::Triple::ArchType::thumb; 1593 discoverJumpTableBranches(MF, isThumb); 1594 } 1595 1596 static bool shouldEmitUdt(const DIType *T) { 1597 if (!T) 1598 return false; 1599 1600 // MSVC does not emit UDTs for typedefs that are scoped to classes. 1601 if (T->getTag() == dwarf::DW_TAG_typedef) { 1602 if (DIScope *Scope = T->getScope()) { 1603 switch (Scope->getTag()) { 1604 case dwarf::DW_TAG_structure_type: 1605 case dwarf::DW_TAG_class_type: 1606 case dwarf::DW_TAG_union_type: 1607 return false; 1608 default: 1609 // do nothing. 1610 ; 1611 } 1612 } 1613 } 1614 1615 while (true) { 1616 if (!T || T->isForwardDecl()) 1617 return false; 1618 1619 const DIDerivedType *DT = dyn_cast<DIDerivedType>(T); 1620 if (!DT) 1621 return true; 1622 T = DT->getBaseType(); 1623 } 1624 return true; 1625 } 1626 1627 void CodeViewDebug::addToUDTs(const DIType *Ty) { 1628 // Don't record empty UDTs. 1629 if (Ty->getName().empty()) 1630 return; 1631 if (!shouldEmitUdt(Ty)) 1632 return; 1633 1634 SmallVector<StringRef, 5> ParentScopeNames; 1635 const DISubprogram *ClosestSubprogram = 1636 collectParentScopeNames(Ty->getScope(), ParentScopeNames); 1637 1638 std::string FullyQualifiedName = 1639 formatNestedName(ParentScopeNames, getPrettyScopeName(Ty)); 1640 1641 if (ClosestSubprogram == nullptr) { 1642 GlobalUDTs.emplace_back(std::move(FullyQualifiedName), Ty); 1643 } else if (ClosestSubprogram == CurrentSubprogram) { 1644 LocalUDTs.emplace_back(std::move(FullyQualifiedName), Ty); 1645 } 1646 1647 // TODO: What if the ClosestSubprogram is neither null or the current 1648 // subprogram? Currently, the UDT just gets dropped on the floor. 1649 // 1650 // The current behavior is not desirable. To get maximal fidelity, we would 1651 // need to perform all type translation before beginning emission of .debug$S 1652 // and then make LocalUDTs a member of FunctionInfo 1653 } 1654 1655 TypeIndex CodeViewDebug::lowerType(const DIType *Ty, const DIType *ClassTy) { 1656 // Generic dispatch for lowering an unknown type. 1657 switch (Ty->getTag()) { 1658 case dwarf::DW_TAG_array_type: 1659 return lowerTypeArray(cast<DICompositeType>(Ty)); 1660 case dwarf::DW_TAG_typedef: 1661 return lowerTypeAlias(cast<DIDerivedType>(Ty)); 1662 case dwarf::DW_TAG_base_type: 1663 return lowerTypeBasic(cast<DIBasicType>(Ty)); 1664 case dwarf::DW_TAG_pointer_type: 1665 if (cast<DIDerivedType>(Ty)->getName() == "__vtbl_ptr_type") 1666 return lowerTypeVFTableShape(cast<DIDerivedType>(Ty)); 1667 [[fallthrough]]; 1668 case dwarf::DW_TAG_reference_type: 1669 case dwarf::DW_TAG_rvalue_reference_type: 1670 return lowerTypePointer(cast<DIDerivedType>(Ty)); 1671 case dwarf::DW_TAG_ptr_to_member_type: 1672 return lowerTypeMemberPointer(cast<DIDerivedType>(Ty)); 1673 case dwarf::DW_TAG_restrict_type: 1674 case dwarf::DW_TAG_const_type: 1675 case dwarf::DW_TAG_volatile_type: 1676 // TODO: add support for DW_TAG_atomic_type here 1677 return lowerTypeModifier(cast<DIDerivedType>(Ty)); 1678 case dwarf::DW_TAG_subroutine_type: 1679 if (ClassTy) { 1680 // The member function type of a member function pointer has no 1681 // ThisAdjustment. 1682 return lowerTypeMemberFunction(cast<DISubroutineType>(Ty), ClassTy, 1683 /*ThisAdjustment=*/0, 1684 /*IsStaticMethod=*/false); 1685 } 1686 return lowerTypeFunction(cast<DISubroutineType>(Ty)); 1687 case dwarf::DW_TAG_enumeration_type: 1688 return lowerTypeEnum(cast<DICompositeType>(Ty)); 1689 case dwarf::DW_TAG_class_type: 1690 case dwarf::DW_TAG_structure_type: 1691 return lowerTypeClass(cast<DICompositeType>(Ty)); 1692 case dwarf::DW_TAG_union_type: 1693 return lowerTypeUnion(cast<DICompositeType>(Ty)); 1694 case dwarf::DW_TAG_string_type: 1695 return lowerTypeString(cast<DIStringType>(Ty)); 1696 case dwarf::DW_TAG_unspecified_type: 1697 if (Ty->getName() == "decltype(nullptr)") 1698 return TypeIndex::NullptrT(); 1699 return TypeIndex::None(); 1700 default: 1701 // Use the null type index. 1702 return TypeIndex(); 1703 } 1704 } 1705 1706 TypeIndex CodeViewDebug::lowerTypeAlias(const DIDerivedType *Ty) { 1707 TypeIndex UnderlyingTypeIndex = getTypeIndex(Ty->getBaseType()); 1708 StringRef TypeName = Ty->getName(); 1709 1710 addToUDTs(Ty); 1711 1712 if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::Int32Long) && 1713 TypeName == "HRESULT") 1714 return TypeIndex(SimpleTypeKind::HResult); 1715 if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::UInt16Short) && 1716 TypeName == "wchar_t") 1717 return TypeIndex(SimpleTypeKind::WideCharacter); 1718 1719 return UnderlyingTypeIndex; 1720 } 1721 1722 TypeIndex CodeViewDebug::lowerTypeArray(const DICompositeType *Ty) { 1723 const DIType *ElementType = Ty->getBaseType(); 1724 TypeIndex ElementTypeIndex = getTypeIndex(ElementType); 1725 // IndexType is size_t, which depends on the bitness of the target. 1726 TypeIndex IndexType = getPointerSizeInBytes() == 8 1727 ? TypeIndex(SimpleTypeKind::UInt64Quad) 1728 : TypeIndex(SimpleTypeKind::UInt32Long); 1729 1730 uint64_t ElementSize = getBaseTypeSize(ElementType) / 8; 1731 1732 // Add subranges to array type. 1733 DINodeArray Elements = Ty->getElements(); 1734 for (int i = Elements.size() - 1; i >= 0; --i) { 1735 const DINode *Element = Elements[i]; 1736 assert(Element->getTag() == dwarf::DW_TAG_subrange_type); 1737 1738 const DISubrange *Subrange = cast<DISubrange>(Element); 1739 int64_t Count = -1; 1740 1741 // If Subrange has a Count field, use it. 1742 // Otherwise, if it has an upperboud, use (upperbound - lowerbound + 1), 1743 // where lowerbound is from the LowerBound field of the Subrange, 1744 // or the language default lowerbound if that field is unspecified. 1745 if (auto *CI = dyn_cast_if_present<ConstantInt *>(Subrange->getCount())) 1746 Count = CI->getSExtValue(); 1747 else if (auto *UI = dyn_cast_if_present<ConstantInt *>( 1748 Subrange->getUpperBound())) { 1749 // Fortran uses 1 as the default lowerbound; other languages use 0. 1750 int64_t Lowerbound = (moduleIsInFortran()) ? 1 : 0; 1751 auto *LI = dyn_cast_if_present<ConstantInt *>(Subrange->getLowerBound()); 1752 Lowerbound = (LI) ? LI->getSExtValue() : Lowerbound; 1753 Count = UI->getSExtValue() - Lowerbound + 1; 1754 } 1755 1756 // Forward declarations of arrays without a size and VLAs use a count of -1. 1757 // Emit a count of zero in these cases to match what MSVC does for arrays 1758 // without a size. MSVC doesn't support VLAs, so it's not clear what we 1759 // should do for them even if we could distinguish them. 1760 if (Count == -1) 1761 Count = 0; 1762 1763 // Update the element size and element type index for subsequent subranges. 1764 ElementSize *= Count; 1765 1766 // If this is the outermost array, use the size from the array. It will be 1767 // more accurate if we had a VLA or an incomplete element type size. 1768 uint64_t ArraySize = 1769 (i == 0 && ElementSize == 0) ? Ty->getSizeInBits() / 8 : ElementSize; 1770 1771 StringRef Name = (i == 0) ? Ty->getName() : ""; 1772 ArrayRecord AR(ElementTypeIndex, IndexType, ArraySize, Name); 1773 ElementTypeIndex = TypeTable.writeLeafType(AR); 1774 } 1775 1776 return ElementTypeIndex; 1777 } 1778 1779 // This function lowers a Fortran character type (DIStringType). 1780 // Note that it handles only the character*n variant (using SizeInBits 1781 // field in DIString to describe the type size) at the moment. 1782 // Other variants (leveraging the StringLength and StringLengthExp 1783 // fields in DIStringType) remain TBD. 1784 TypeIndex CodeViewDebug::lowerTypeString(const DIStringType *Ty) { 1785 TypeIndex CharType = TypeIndex(SimpleTypeKind::NarrowCharacter); 1786 uint64_t ArraySize = Ty->getSizeInBits() >> 3; 1787 StringRef Name = Ty->getName(); 1788 // IndexType is size_t, which depends on the bitness of the target. 1789 TypeIndex IndexType = getPointerSizeInBytes() == 8 1790 ? TypeIndex(SimpleTypeKind::UInt64Quad) 1791 : TypeIndex(SimpleTypeKind::UInt32Long); 1792 1793 // Create a type of character array of ArraySize. 1794 ArrayRecord AR(CharType, IndexType, ArraySize, Name); 1795 1796 return TypeTable.writeLeafType(AR); 1797 } 1798 1799 TypeIndex CodeViewDebug::lowerTypeBasic(const DIBasicType *Ty) { 1800 TypeIndex Index; 1801 dwarf::TypeKind Kind; 1802 uint32_t ByteSize; 1803 1804 Kind = static_cast<dwarf::TypeKind>(Ty->getEncoding()); 1805 ByteSize = Ty->getSizeInBits() / 8; 1806 1807 SimpleTypeKind STK = SimpleTypeKind::None; 1808 switch (Kind) { 1809 case dwarf::DW_ATE_address: 1810 // FIXME: Translate 1811 break; 1812 case dwarf::DW_ATE_boolean: 1813 switch (ByteSize) { 1814 case 1: STK = SimpleTypeKind::Boolean8; break; 1815 case 2: STK = SimpleTypeKind::Boolean16; break; 1816 case 4: STK = SimpleTypeKind::Boolean32; break; 1817 case 8: STK = SimpleTypeKind::Boolean64; break; 1818 case 16: STK = SimpleTypeKind::Boolean128; break; 1819 } 1820 break; 1821 case dwarf::DW_ATE_complex_float: 1822 // The CodeView size for a complex represents the size of 1823 // an individual component. 1824 switch (ByteSize) { 1825 case 4: STK = SimpleTypeKind::Complex16; break; 1826 case 8: STK = SimpleTypeKind::Complex32; break; 1827 case 16: STK = SimpleTypeKind::Complex64; break; 1828 case 20: STK = SimpleTypeKind::Complex80; break; 1829 case 32: STK = SimpleTypeKind::Complex128; break; 1830 } 1831 break; 1832 case dwarf::DW_ATE_float: 1833 switch (ByteSize) { 1834 case 2: STK = SimpleTypeKind::Float16; break; 1835 case 4: STK = SimpleTypeKind::Float32; break; 1836 case 6: STK = SimpleTypeKind::Float48; break; 1837 case 8: STK = SimpleTypeKind::Float64; break; 1838 case 10: STK = SimpleTypeKind::Float80; break; 1839 case 16: STK = SimpleTypeKind::Float128; break; 1840 } 1841 break; 1842 case dwarf::DW_ATE_signed: 1843 switch (ByteSize) { 1844 case 1: STK = SimpleTypeKind::SignedCharacter; break; 1845 case 2: STK = SimpleTypeKind::Int16Short; break; 1846 case 4: STK = SimpleTypeKind::Int32; break; 1847 case 8: STK = SimpleTypeKind::Int64Quad; break; 1848 case 16: STK = SimpleTypeKind::Int128Oct; break; 1849 } 1850 break; 1851 case dwarf::DW_ATE_unsigned: 1852 switch (ByteSize) { 1853 case 1: STK = SimpleTypeKind::UnsignedCharacter; break; 1854 case 2: STK = SimpleTypeKind::UInt16Short; break; 1855 case 4: STK = SimpleTypeKind::UInt32; break; 1856 case 8: STK = SimpleTypeKind::UInt64Quad; break; 1857 case 16: STK = SimpleTypeKind::UInt128Oct; break; 1858 } 1859 break; 1860 case dwarf::DW_ATE_UTF: 1861 switch (ByteSize) { 1862 case 1: STK = SimpleTypeKind::Character8; break; 1863 case 2: STK = SimpleTypeKind::Character16; break; 1864 case 4: STK = SimpleTypeKind::Character32; break; 1865 } 1866 break; 1867 case dwarf::DW_ATE_signed_char: 1868 if (ByteSize == 1) 1869 STK = SimpleTypeKind::SignedCharacter; 1870 break; 1871 case dwarf::DW_ATE_unsigned_char: 1872 if (ByteSize == 1) 1873 STK = SimpleTypeKind::UnsignedCharacter; 1874 break; 1875 default: 1876 break; 1877 } 1878 1879 // Apply some fixups based on the source-level type name. 1880 // Include some amount of canonicalization from an old naming scheme Clang 1881 // used to use for integer types (in an outdated effort to be compatible with 1882 // GCC's debug info/GDB's behavior, which has since been addressed). 1883 if (STK == SimpleTypeKind::Int32 && 1884 (Ty->getName() == "long int" || Ty->getName() == "long")) 1885 STK = SimpleTypeKind::Int32Long; 1886 if (STK == SimpleTypeKind::UInt32 && (Ty->getName() == "long unsigned int" || 1887 Ty->getName() == "unsigned long")) 1888 STK = SimpleTypeKind::UInt32Long; 1889 if (STK == SimpleTypeKind::UInt16Short && 1890 (Ty->getName() == "wchar_t" || Ty->getName() == "__wchar_t")) 1891 STK = SimpleTypeKind::WideCharacter; 1892 if ((STK == SimpleTypeKind::SignedCharacter || 1893 STK == SimpleTypeKind::UnsignedCharacter) && 1894 Ty->getName() == "char") 1895 STK = SimpleTypeKind::NarrowCharacter; 1896 1897 return TypeIndex(STK); 1898 } 1899 1900 TypeIndex CodeViewDebug::lowerTypePointer(const DIDerivedType *Ty, 1901 PointerOptions PO) { 1902 TypeIndex PointeeTI = getTypeIndex(Ty->getBaseType()); 1903 1904 // Pointers to simple types without any options can use SimpleTypeMode, rather 1905 // than having a dedicated pointer type record. 1906 if (PointeeTI.isSimple() && PO == PointerOptions::None && 1907 PointeeTI.getSimpleMode() == SimpleTypeMode::Direct && 1908 Ty->getTag() == dwarf::DW_TAG_pointer_type) { 1909 SimpleTypeMode Mode = Ty->getSizeInBits() == 64 1910 ? SimpleTypeMode::NearPointer64 1911 : SimpleTypeMode::NearPointer32; 1912 return TypeIndex(PointeeTI.getSimpleKind(), Mode); 1913 } 1914 1915 PointerKind PK = 1916 Ty->getSizeInBits() == 64 ? PointerKind::Near64 : PointerKind::Near32; 1917 PointerMode PM = PointerMode::Pointer; 1918 switch (Ty->getTag()) { 1919 default: llvm_unreachable("not a pointer tag type"); 1920 case dwarf::DW_TAG_pointer_type: 1921 PM = PointerMode::Pointer; 1922 break; 1923 case dwarf::DW_TAG_reference_type: 1924 PM = PointerMode::LValueReference; 1925 break; 1926 case dwarf::DW_TAG_rvalue_reference_type: 1927 PM = PointerMode::RValueReference; 1928 break; 1929 } 1930 1931 if (Ty->isObjectPointer()) 1932 PO |= PointerOptions::Const; 1933 1934 PointerRecord PR(PointeeTI, PK, PM, PO, Ty->getSizeInBits() / 8); 1935 return TypeTable.writeLeafType(PR); 1936 } 1937 1938 static PointerToMemberRepresentation 1939 translatePtrToMemberRep(unsigned SizeInBytes, bool IsPMF, unsigned Flags) { 1940 // SizeInBytes being zero generally implies that the member pointer type was 1941 // incomplete, which can happen if it is part of a function prototype. In this 1942 // case, use the unknown model instead of the general model. 1943 if (IsPMF) { 1944 switch (Flags & DINode::FlagPtrToMemberRep) { 1945 case 0: 1946 return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown 1947 : PointerToMemberRepresentation::GeneralFunction; 1948 case DINode::FlagSingleInheritance: 1949 return PointerToMemberRepresentation::SingleInheritanceFunction; 1950 case DINode::FlagMultipleInheritance: 1951 return PointerToMemberRepresentation::MultipleInheritanceFunction; 1952 case DINode::FlagVirtualInheritance: 1953 return PointerToMemberRepresentation::VirtualInheritanceFunction; 1954 } 1955 } else { 1956 switch (Flags & DINode::FlagPtrToMemberRep) { 1957 case 0: 1958 return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown 1959 : PointerToMemberRepresentation::GeneralData; 1960 case DINode::FlagSingleInheritance: 1961 return PointerToMemberRepresentation::SingleInheritanceData; 1962 case DINode::FlagMultipleInheritance: 1963 return PointerToMemberRepresentation::MultipleInheritanceData; 1964 case DINode::FlagVirtualInheritance: 1965 return PointerToMemberRepresentation::VirtualInheritanceData; 1966 } 1967 } 1968 llvm_unreachable("invalid ptr to member representation"); 1969 } 1970 1971 TypeIndex CodeViewDebug::lowerTypeMemberPointer(const DIDerivedType *Ty, 1972 PointerOptions PO) { 1973 assert(Ty->getTag() == dwarf::DW_TAG_ptr_to_member_type); 1974 bool IsPMF = isa<DISubroutineType>(Ty->getBaseType()); 1975 TypeIndex ClassTI = getTypeIndex(Ty->getClassType()); 1976 TypeIndex PointeeTI = 1977 getTypeIndex(Ty->getBaseType(), IsPMF ? Ty->getClassType() : nullptr); 1978 PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64 1979 : PointerKind::Near32; 1980 PointerMode PM = IsPMF ? PointerMode::PointerToMemberFunction 1981 : PointerMode::PointerToDataMember; 1982 1983 assert(Ty->getSizeInBits() / 8 <= 0xff && "pointer size too big"); 1984 uint8_t SizeInBytes = Ty->getSizeInBits() / 8; 1985 MemberPointerInfo MPI( 1986 ClassTI, translatePtrToMemberRep(SizeInBytes, IsPMF, Ty->getFlags())); 1987 PointerRecord PR(PointeeTI, PK, PM, PO, SizeInBytes, MPI); 1988 return TypeTable.writeLeafType(PR); 1989 } 1990 1991 /// Given a DWARF calling convention, get the CodeView equivalent. If we don't 1992 /// have a translation, use the NearC convention. 1993 static CallingConvention dwarfCCToCodeView(unsigned DwarfCC) { 1994 switch (DwarfCC) { 1995 case dwarf::DW_CC_normal: return CallingConvention::NearC; 1996 case dwarf::DW_CC_BORLAND_msfastcall: return CallingConvention::NearFast; 1997 case dwarf::DW_CC_BORLAND_thiscall: return CallingConvention::ThisCall; 1998 case dwarf::DW_CC_BORLAND_stdcall: return CallingConvention::NearStdCall; 1999 case dwarf::DW_CC_BORLAND_pascal: return CallingConvention::NearPascal; 2000 case dwarf::DW_CC_LLVM_vectorcall: return CallingConvention::NearVector; 2001 } 2002 return CallingConvention::NearC; 2003 } 2004 2005 TypeIndex CodeViewDebug::lowerTypeModifier(const DIDerivedType *Ty) { 2006 ModifierOptions Mods = ModifierOptions::None; 2007 PointerOptions PO = PointerOptions::None; 2008 bool IsModifier = true; 2009 const DIType *BaseTy = Ty; 2010 while (IsModifier && BaseTy) { 2011 // FIXME: Need to add DWARF tags for __unaligned and _Atomic 2012 switch (BaseTy->getTag()) { 2013 case dwarf::DW_TAG_const_type: 2014 Mods |= ModifierOptions::Const; 2015 PO |= PointerOptions::Const; 2016 break; 2017 case dwarf::DW_TAG_volatile_type: 2018 Mods |= ModifierOptions::Volatile; 2019 PO |= PointerOptions::Volatile; 2020 break; 2021 case dwarf::DW_TAG_restrict_type: 2022 // Only pointer types be marked with __restrict. There is no known flag 2023 // for __restrict in LF_MODIFIER records. 2024 PO |= PointerOptions::Restrict; 2025 break; 2026 default: 2027 IsModifier = false; 2028 break; 2029 } 2030 if (IsModifier) 2031 BaseTy = cast<DIDerivedType>(BaseTy)->getBaseType(); 2032 } 2033 2034 // Check if the inner type will use an LF_POINTER record. If so, the 2035 // qualifiers will go in the LF_POINTER record. This comes up for types like 2036 // 'int *const' and 'int *__restrict', not the more common cases like 'const 2037 // char *'. 2038 if (BaseTy) { 2039 switch (BaseTy->getTag()) { 2040 case dwarf::DW_TAG_pointer_type: 2041 case dwarf::DW_TAG_reference_type: 2042 case dwarf::DW_TAG_rvalue_reference_type: 2043 return lowerTypePointer(cast<DIDerivedType>(BaseTy), PO); 2044 case dwarf::DW_TAG_ptr_to_member_type: 2045 return lowerTypeMemberPointer(cast<DIDerivedType>(BaseTy), PO); 2046 default: 2047 break; 2048 } 2049 } 2050 2051 TypeIndex ModifiedTI = getTypeIndex(BaseTy); 2052 2053 // Return the base type index if there aren't any modifiers. For example, the 2054 // metadata could contain restrict wrappers around non-pointer types. 2055 if (Mods == ModifierOptions::None) 2056 return ModifiedTI; 2057 2058 ModifierRecord MR(ModifiedTI, Mods); 2059 return TypeTable.writeLeafType(MR); 2060 } 2061 2062 TypeIndex CodeViewDebug::lowerTypeFunction(const DISubroutineType *Ty) { 2063 SmallVector<TypeIndex, 8> ReturnAndArgTypeIndices; 2064 for (const DIType *ArgType : Ty->getTypeArray()) 2065 ReturnAndArgTypeIndices.push_back(getTypeIndex(ArgType)); 2066 2067 // MSVC uses type none for variadic argument. 2068 if (ReturnAndArgTypeIndices.size() > 1 && 2069 ReturnAndArgTypeIndices.back() == TypeIndex::Void()) { 2070 ReturnAndArgTypeIndices.back() = TypeIndex::None(); 2071 } 2072 TypeIndex ReturnTypeIndex = TypeIndex::Void(); 2073 ArrayRef<TypeIndex> ArgTypeIndices = std::nullopt; 2074 if (!ReturnAndArgTypeIndices.empty()) { 2075 auto ReturnAndArgTypesRef = ArrayRef(ReturnAndArgTypeIndices); 2076 ReturnTypeIndex = ReturnAndArgTypesRef.front(); 2077 ArgTypeIndices = ReturnAndArgTypesRef.drop_front(); 2078 } 2079 2080 ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices); 2081 TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec); 2082 2083 CallingConvention CC = dwarfCCToCodeView(Ty->getCC()); 2084 2085 FunctionOptions FO = getFunctionOptions(Ty); 2086 ProcedureRecord Procedure(ReturnTypeIndex, CC, FO, ArgTypeIndices.size(), 2087 ArgListIndex); 2088 return TypeTable.writeLeafType(Procedure); 2089 } 2090 2091 TypeIndex CodeViewDebug::lowerTypeMemberFunction(const DISubroutineType *Ty, 2092 const DIType *ClassTy, 2093 int ThisAdjustment, 2094 bool IsStaticMethod, 2095 FunctionOptions FO) { 2096 // Lower the containing class type. 2097 TypeIndex ClassType = getTypeIndex(ClassTy); 2098 2099 DITypeRefArray ReturnAndArgs = Ty->getTypeArray(); 2100 2101 unsigned Index = 0; 2102 SmallVector<TypeIndex, 8> ArgTypeIndices; 2103 TypeIndex ReturnTypeIndex = TypeIndex::Void(); 2104 if (ReturnAndArgs.size() > Index) { 2105 ReturnTypeIndex = getTypeIndex(ReturnAndArgs[Index++]); 2106 } 2107 2108 // If the first argument is a pointer type and this isn't a static method, 2109 // treat it as the special 'this' parameter, which is encoded separately from 2110 // the arguments. 2111 TypeIndex ThisTypeIndex; 2112 if (!IsStaticMethod && ReturnAndArgs.size() > Index) { 2113 if (const DIDerivedType *PtrTy = 2114 dyn_cast_or_null<DIDerivedType>(ReturnAndArgs[Index])) { 2115 if (PtrTy->getTag() == dwarf::DW_TAG_pointer_type) { 2116 ThisTypeIndex = getTypeIndexForThisPtr(PtrTy, Ty); 2117 Index++; 2118 } 2119 } 2120 } 2121 2122 while (Index < ReturnAndArgs.size()) 2123 ArgTypeIndices.push_back(getTypeIndex(ReturnAndArgs[Index++])); 2124 2125 // MSVC uses type none for variadic argument. 2126 if (!ArgTypeIndices.empty() && ArgTypeIndices.back() == TypeIndex::Void()) 2127 ArgTypeIndices.back() = TypeIndex::None(); 2128 2129 ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices); 2130 TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec); 2131 2132 CallingConvention CC = dwarfCCToCodeView(Ty->getCC()); 2133 2134 MemberFunctionRecord MFR(ReturnTypeIndex, ClassType, ThisTypeIndex, CC, FO, 2135 ArgTypeIndices.size(), ArgListIndex, ThisAdjustment); 2136 return TypeTable.writeLeafType(MFR); 2137 } 2138 2139 TypeIndex CodeViewDebug::lowerTypeVFTableShape(const DIDerivedType *Ty) { 2140 unsigned VSlotCount = 2141 Ty->getSizeInBits() / (8 * Asm->MAI->getCodePointerSize()); 2142 SmallVector<VFTableSlotKind, 4> Slots(VSlotCount, VFTableSlotKind::Near); 2143 2144 VFTableShapeRecord VFTSR(Slots); 2145 return TypeTable.writeLeafType(VFTSR); 2146 } 2147 2148 static MemberAccess translateAccessFlags(unsigned RecordTag, unsigned Flags) { 2149 switch (Flags & DINode::FlagAccessibility) { 2150 case DINode::FlagPrivate: return MemberAccess::Private; 2151 case DINode::FlagPublic: return MemberAccess::Public; 2152 case DINode::FlagProtected: return MemberAccess::Protected; 2153 case 0: 2154 // If there was no explicit access control, provide the default for the tag. 2155 return RecordTag == dwarf::DW_TAG_class_type ? MemberAccess::Private 2156 : MemberAccess::Public; 2157 } 2158 llvm_unreachable("access flags are exclusive"); 2159 } 2160 2161 static MethodOptions translateMethodOptionFlags(const DISubprogram *SP) { 2162 if (SP->isArtificial()) 2163 return MethodOptions::CompilerGenerated; 2164 2165 // FIXME: Handle other MethodOptions. 2166 2167 return MethodOptions::None; 2168 } 2169 2170 static MethodKind translateMethodKindFlags(const DISubprogram *SP, 2171 bool Introduced) { 2172 if (SP->getFlags() & DINode::FlagStaticMember) 2173 return MethodKind::Static; 2174 2175 switch (SP->getVirtuality()) { 2176 case dwarf::DW_VIRTUALITY_none: 2177 break; 2178 case dwarf::DW_VIRTUALITY_virtual: 2179 return Introduced ? MethodKind::IntroducingVirtual : MethodKind::Virtual; 2180 case dwarf::DW_VIRTUALITY_pure_virtual: 2181 return Introduced ? MethodKind::PureIntroducingVirtual 2182 : MethodKind::PureVirtual; 2183 default: 2184 llvm_unreachable("unhandled virtuality case"); 2185 } 2186 2187 return MethodKind::Vanilla; 2188 } 2189 2190 static TypeRecordKind getRecordKind(const DICompositeType *Ty) { 2191 switch (Ty->getTag()) { 2192 case dwarf::DW_TAG_class_type: 2193 return TypeRecordKind::Class; 2194 case dwarf::DW_TAG_structure_type: 2195 return TypeRecordKind::Struct; 2196 default: 2197 llvm_unreachable("unexpected tag"); 2198 } 2199 } 2200 2201 /// Return ClassOptions that should be present on both the forward declaration 2202 /// and the defintion of a tag type. 2203 static ClassOptions getCommonClassOptions(const DICompositeType *Ty) { 2204 ClassOptions CO = ClassOptions::None; 2205 2206 // MSVC always sets this flag, even for local types. Clang doesn't always 2207 // appear to give every type a linkage name, which may be problematic for us. 2208 // FIXME: Investigate the consequences of not following them here. 2209 if (!Ty->getIdentifier().empty()) 2210 CO |= ClassOptions::HasUniqueName; 2211 2212 // Put the Nested flag on a type if it appears immediately inside a tag type. 2213 // Do not walk the scope chain. Do not attempt to compute ContainsNestedClass 2214 // here. That flag is only set on definitions, and not forward declarations. 2215 const DIScope *ImmediateScope = Ty->getScope(); 2216 if (ImmediateScope && isa<DICompositeType>(ImmediateScope)) 2217 CO |= ClassOptions::Nested; 2218 2219 // Put the Scoped flag on function-local types. MSVC puts this flag for enum 2220 // type only when it has an immediate function scope. Clang never puts enums 2221 // inside DILexicalBlock scopes. Enum types, as generated by clang, are 2222 // always in function, class, or file scopes. 2223 if (Ty->getTag() == dwarf::DW_TAG_enumeration_type) { 2224 if (ImmediateScope && isa<DISubprogram>(ImmediateScope)) 2225 CO |= ClassOptions::Scoped; 2226 } else { 2227 for (const DIScope *Scope = ImmediateScope; Scope != nullptr; 2228 Scope = Scope->getScope()) { 2229 if (isa<DISubprogram>(Scope)) { 2230 CO |= ClassOptions::Scoped; 2231 break; 2232 } 2233 } 2234 } 2235 2236 return CO; 2237 } 2238 2239 void CodeViewDebug::addUDTSrcLine(const DIType *Ty, TypeIndex TI) { 2240 switch (Ty->getTag()) { 2241 case dwarf::DW_TAG_class_type: 2242 case dwarf::DW_TAG_structure_type: 2243 case dwarf::DW_TAG_union_type: 2244 case dwarf::DW_TAG_enumeration_type: 2245 break; 2246 default: 2247 return; 2248 } 2249 2250 if (const auto *File = Ty->getFile()) { 2251 StringIdRecord SIDR(TypeIndex(0x0), getFullFilepath(File)); 2252 TypeIndex SIDI = TypeTable.writeLeafType(SIDR); 2253 2254 UdtSourceLineRecord USLR(TI, SIDI, Ty->getLine()); 2255 TypeTable.writeLeafType(USLR); 2256 } 2257 } 2258 2259 TypeIndex CodeViewDebug::lowerTypeEnum(const DICompositeType *Ty) { 2260 ClassOptions CO = getCommonClassOptions(Ty); 2261 TypeIndex FTI; 2262 unsigned EnumeratorCount = 0; 2263 2264 if (Ty->isForwardDecl()) { 2265 CO |= ClassOptions::ForwardReference; 2266 } else { 2267 ContinuationRecordBuilder ContinuationBuilder; 2268 ContinuationBuilder.begin(ContinuationRecordKind::FieldList); 2269 for (const DINode *Element : Ty->getElements()) { 2270 // We assume that the frontend provides all members in source declaration 2271 // order, which is what MSVC does. 2272 if (auto *Enumerator = dyn_cast_or_null<DIEnumerator>(Element)) { 2273 // FIXME: Is it correct to always emit these as unsigned here? 2274 EnumeratorRecord ER(MemberAccess::Public, 2275 APSInt(Enumerator->getValue(), true), 2276 Enumerator->getName()); 2277 ContinuationBuilder.writeMemberType(ER); 2278 EnumeratorCount++; 2279 } 2280 } 2281 FTI = TypeTable.insertRecord(ContinuationBuilder); 2282 } 2283 2284 std::string FullName = getFullyQualifiedName(Ty); 2285 2286 EnumRecord ER(EnumeratorCount, CO, FTI, FullName, Ty->getIdentifier(), 2287 getTypeIndex(Ty->getBaseType())); 2288 TypeIndex EnumTI = TypeTable.writeLeafType(ER); 2289 2290 addUDTSrcLine(Ty, EnumTI); 2291 2292 return EnumTI; 2293 } 2294 2295 //===----------------------------------------------------------------------===// 2296 // ClassInfo 2297 //===----------------------------------------------------------------------===// 2298 2299 struct llvm::ClassInfo { 2300 struct MemberInfo { 2301 const DIDerivedType *MemberTypeNode; 2302 uint64_t BaseOffset; 2303 }; 2304 // [MemberInfo] 2305 using MemberList = std::vector<MemberInfo>; 2306 2307 using MethodsList = TinyPtrVector<const DISubprogram *>; 2308 // MethodName -> MethodsList 2309 using MethodsMap = MapVector<MDString *, MethodsList>; 2310 2311 /// Base classes. 2312 std::vector<const DIDerivedType *> Inheritance; 2313 2314 /// Direct members. 2315 MemberList Members; 2316 // Direct overloaded methods gathered by name. 2317 MethodsMap Methods; 2318 2319 TypeIndex VShapeTI; 2320 2321 std::vector<const DIType *> NestedTypes; 2322 }; 2323 2324 void CodeViewDebug::clear() { 2325 assert(CurFn == nullptr); 2326 FileIdMap.clear(); 2327 FnDebugInfo.clear(); 2328 FileToFilepathMap.clear(); 2329 LocalUDTs.clear(); 2330 GlobalUDTs.clear(); 2331 TypeIndices.clear(); 2332 CompleteTypeIndices.clear(); 2333 ScopeGlobals.clear(); 2334 CVGlobalVariableOffsets.clear(); 2335 } 2336 2337 void CodeViewDebug::collectMemberInfo(ClassInfo &Info, 2338 const DIDerivedType *DDTy) { 2339 if (!DDTy->getName().empty()) { 2340 Info.Members.push_back({DDTy, 0}); 2341 2342 // Collect static const data members with values. 2343 if ((DDTy->getFlags() & DINode::FlagStaticMember) == 2344 DINode::FlagStaticMember) { 2345 if (DDTy->getConstant() && (isa<ConstantInt>(DDTy->getConstant()) || 2346 isa<ConstantFP>(DDTy->getConstant()))) 2347 StaticConstMembers.push_back(DDTy); 2348 } 2349 2350 return; 2351 } 2352 2353 // An unnamed member may represent a nested struct or union. Attempt to 2354 // interpret the unnamed member as a DICompositeType possibly wrapped in 2355 // qualifier types. Add all the indirect fields to the current record if that 2356 // succeeds, and drop the member if that fails. 2357 assert((DDTy->getOffsetInBits() % 8) == 0 && "Unnamed bitfield member!"); 2358 uint64_t Offset = DDTy->getOffsetInBits(); 2359 const DIType *Ty = DDTy->getBaseType(); 2360 bool FullyResolved = false; 2361 while (!FullyResolved) { 2362 switch (Ty->getTag()) { 2363 case dwarf::DW_TAG_const_type: 2364 case dwarf::DW_TAG_volatile_type: 2365 // FIXME: we should apply the qualifier types to the indirect fields 2366 // rather than dropping them. 2367 Ty = cast<DIDerivedType>(Ty)->getBaseType(); 2368 break; 2369 default: 2370 FullyResolved = true; 2371 break; 2372 } 2373 } 2374 2375 const DICompositeType *DCTy = dyn_cast<DICompositeType>(Ty); 2376 if (!DCTy) 2377 return; 2378 2379 ClassInfo NestedInfo = collectClassInfo(DCTy); 2380 for (const ClassInfo::MemberInfo &IndirectField : NestedInfo.Members) 2381 Info.Members.push_back( 2382 {IndirectField.MemberTypeNode, IndirectField.BaseOffset + Offset}); 2383 } 2384 2385 ClassInfo CodeViewDebug::collectClassInfo(const DICompositeType *Ty) { 2386 ClassInfo Info; 2387 // Add elements to structure type. 2388 DINodeArray Elements = Ty->getElements(); 2389 for (auto *Element : Elements) { 2390 // We assume that the frontend provides all members in source declaration 2391 // order, which is what MSVC does. 2392 if (!Element) 2393 continue; 2394 if (auto *SP = dyn_cast<DISubprogram>(Element)) { 2395 Info.Methods[SP->getRawName()].push_back(SP); 2396 } else if (auto *DDTy = dyn_cast<DIDerivedType>(Element)) { 2397 if (DDTy->getTag() == dwarf::DW_TAG_member) { 2398 collectMemberInfo(Info, DDTy); 2399 } else if (DDTy->getTag() == dwarf::DW_TAG_inheritance) { 2400 Info.Inheritance.push_back(DDTy); 2401 } else if (DDTy->getTag() == dwarf::DW_TAG_pointer_type && 2402 DDTy->getName() == "__vtbl_ptr_type") { 2403 Info.VShapeTI = getTypeIndex(DDTy); 2404 } else if (DDTy->getTag() == dwarf::DW_TAG_typedef) { 2405 Info.NestedTypes.push_back(DDTy); 2406 } else if (DDTy->getTag() == dwarf::DW_TAG_friend) { 2407 // Ignore friend members. It appears that MSVC emitted info about 2408 // friends in the past, but modern versions do not. 2409 } 2410 } else if (auto *Composite = dyn_cast<DICompositeType>(Element)) { 2411 Info.NestedTypes.push_back(Composite); 2412 } 2413 // Skip other unrecognized kinds of elements. 2414 } 2415 return Info; 2416 } 2417 2418 static bool shouldAlwaysEmitCompleteClassType(const DICompositeType *Ty) { 2419 // This routine is used by lowerTypeClass and lowerTypeUnion to determine 2420 // if a complete type should be emitted instead of a forward reference. 2421 return Ty->getName().empty() && Ty->getIdentifier().empty() && 2422 !Ty->isForwardDecl(); 2423 } 2424 2425 TypeIndex CodeViewDebug::lowerTypeClass(const DICompositeType *Ty) { 2426 // Emit the complete type for unnamed structs. C++ classes with methods 2427 // which have a circular reference back to the class type are expected to 2428 // be named by the front-end and should not be "unnamed". C unnamed 2429 // structs should not have circular references. 2430 if (shouldAlwaysEmitCompleteClassType(Ty)) { 2431 // If this unnamed complete type is already in the process of being defined 2432 // then the description of the type is malformed and cannot be emitted 2433 // into CodeView correctly so report a fatal error. 2434 auto I = CompleteTypeIndices.find(Ty); 2435 if (I != CompleteTypeIndices.end() && I->second == TypeIndex()) 2436 report_fatal_error("cannot debug circular reference to unnamed type"); 2437 return getCompleteTypeIndex(Ty); 2438 } 2439 2440 // First, construct the forward decl. Don't look into Ty to compute the 2441 // forward decl options, since it might not be available in all TUs. 2442 TypeRecordKind Kind = getRecordKind(Ty); 2443 ClassOptions CO = 2444 ClassOptions::ForwardReference | getCommonClassOptions(Ty); 2445 std::string FullName = getFullyQualifiedName(Ty); 2446 ClassRecord CR(Kind, 0, CO, TypeIndex(), TypeIndex(), TypeIndex(), 0, 2447 FullName, Ty->getIdentifier()); 2448 TypeIndex FwdDeclTI = TypeTable.writeLeafType(CR); 2449 if (!Ty->isForwardDecl()) 2450 DeferredCompleteTypes.push_back(Ty); 2451 return FwdDeclTI; 2452 } 2453 2454 TypeIndex CodeViewDebug::lowerCompleteTypeClass(const DICompositeType *Ty) { 2455 // Construct the field list and complete type record. 2456 TypeRecordKind Kind = getRecordKind(Ty); 2457 ClassOptions CO = getCommonClassOptions(Ty); 2458 TypeIndex FieldTI; 2459 TypeIndex VShapeTI; 2460 unsigned FieldCount; 2461 bool ContainsNestedClass; 2462 std::tie(FieldTI, VShapeTI, FieldCount, ContainsNestedClass) = 2463 lowerRecordFieldList(Ty); 2464 2465 if (ContainsNestedClass) 2466 CO |= ClassOptions::ContainsNestedClass; 2467 2468 // MSVC appears to set this flag by searching any destructor or method with 2469 // FunctionOptions::Constructor among the emitted members. Clang AST has all 2470 // the members, however special member functions are not yet emitted into 2471 // debug information. For now checking a class's non-triviality seems enough. 2472 // FIXME: not true for a nested unnamed struct. 2473 if (isNonTrivial(Ty)) 2474 CO |= ClassOptions::HasConstructorOrDestructor; 2475 2476 std::string FullName = getFullyQualifiedName(Ty); 2477 2478 uint64_t SizeInBytes = Ty->getSizeInBits() / 8; 2479 2480 ClassRecord CR(Kind, FieldCount, CO, FieldTI, TypeIndex(), VShapeTI, 2481 SizeInBytes, FullName, Ty->getIdentifier()); 2482 TypeIndex ClassTI = TypeTable.writeLeafType(CR); 2483 2484 addUDTSrcLine(Ty, ClassTI); 2485 2486 addToUDTs(Ty); 2487 2488 return ClassTI; 2489 } 2490 2491 TypeIndex CodeViewDebug::lowerTypeUnion(const DICompositeType *Ty) { 2492 // Emit the complete type for unnamed unions. 2493 if (shouldAlwaysEmitCompleteClassType(Ty)) 2494 return getCompleteTypeIndex(Ty); 2495 2496 ClassOptions CO = 2497 ClassOptions::ForwardReference | getCommonClassOptions(Ty); 2498 std::string FullName = getFullyQualifiedName(Ty); 2499 UnionRecord UR(0, CO, TypeIndex(), 0, FullName, Ty->getIdentifier()); 2500 TypeIndex FwdDeclTI = TypeTable.writeLeafType(UR); 2501 if (!Ty->isForwardDecl()) 2502 DeferredCompleteTypes.push_back(Ty); 2503 return FwdDeclTI; 2504 } 2505 2506 TypeIndex CodeViewDebug::lowerCompleteTypeUnion(const DICompositeType *Ty) { 2507 ClassOptions CO = ClassOptions::Sealed | getCommonClassOptions(Ty); 2508 TypeIndex FieldTI; 2509 unsigned FieldCount; 2510 bool ContainsNestedClass; 2511 std::tie(FieldTI, std::ignore, FieldCount, ContainsNestedClass) = 2512 lowerRecordFieldList(Ty); 2513 2514 if (ContainsNestedClass) 2515 CO |= ClassOptions::ContainsNestedClass; 2516 2517 uint64_t SizeInBytes = Ty->getSizeInBits() / 8; 2518 std::string FullName = getFullyQualifiedName(Ty); 2519 2520 UnionRecord UR(FieldCount, CO, FieldTI, SizeInBytes, FullName, 2521 Ty->getIdentifier()); 2522 TypeIndex UnionTI = TypeTable.writeLeafType(UR); 2523 2524 addUDTSrcLine(Ty, UnionTI); 2525 2526 addToUDTs(Ty); 2527 2528 return UnionTI; 2529 } 2530 2531 std::tuple<TypeIndex, TypeIndex, unsigned, bool> 2532 CodeViewDebug::lowerRecordFieldList(const DICompositeType *Ty) { 2533 // Manually count members. MSVC appears to count everything that generates a 2534 // field list record. Each individual overload in a method overload group 2535 // contributes to this count, even though the overload group is a single field 2536 // list record. 2537 unsigned MemberCount = 0; 2538 ClassInfo Info = collectClassInfo(Ty); 2539 ContinuationRecordBuilder ContinuationBuilder; 2540 ContinuationBuilder.begin(ContinuationRecordKind::FieldList); 2541 2542 // Create base classes. 2543 for (const DIDerivedType *I : Info.Inheritance) { 2544 if (I->getFlags() & DINode::FlagVirtual) { 2545 // Virtual base. 2546 unsigned VBPtrOffset = I->getVBPtrOffset(); 2547 // FIXME: Despite the accessor name, the offset is really in bytes. 2548 unsigned VBTableIndex = I->getOffsetInBits() / 4; 2549 auto RecordKind = (I->getFlags() & DINode::FlagIndirectVirtualBase) == DINode::FlagIndirectVirtualBase 2550 ? TypeRecordKind::IndirectVirtualBaseClass 2551 : TypeRecordKind::VirtualBaseClass; 2552 VirtualBaseClassRecord VBCR( 2553 RecordKind, translateAccessFlags(Ty->getTag(), I->getFlags()), 2554 getTypeIndex(I->getBaseType()), getVBPTypeIndex(), VBPtrOffset, 2555 VBTableIndex); 2556 2557 ContinuationBuilder.writeMemberType(VBCR); 2558 MemberCount++; 2559 } else { 2560 assert(I->getOffsetInBits() % 8 == 0 && 2561 "bases must be on byte boundaries"); 2562 BaseClassRecord BCR(translateAccessFlags(Ty->getTag(), I->getFlags()), 2563 getTypeIndex(I->getBaseType()), 2564 I->getOffsetInBits() / 8); 2565 ContinuationBuilder.writeMemberType(BCR); 2566 MemberCount++; 2567 } 2568 } 2569 2570 // Create members. 2571 for (ClassInfo::MemberInfo &MemberInfo : Info.Members) { 2572 const DIDerivedType *Member = MemberInfo.MemberTypeNode; 2573 TypeIndex MemberBaseType = getTypeIndex(Member->getBaseType()); 2574 StringRef MemberName = Member->getName(); 2575 MemberAccess Access = 2576 translateAccessFlags(Ty->getTag(), Member->getFlags()); 2577 2578 if (Member->isStaticMember()) { 2579 StaticDataMemberRecord SDMR(Access, MemberBaseType, MemberName); 2580 ContinuationBuilder.writeMemberType(SDMR); 2581 MemberCount++; 2582 continue; 2583 } 2584 2585 // Virtual function pointer member. 2586 if ((Member->getFlags() & DINode::FlagArtificial) && 2587 Member->getName().startswith("_vptr$")) { 2588 VFPtrRecord VFPR(getTypeIndex(Member->getBaseType())); 2589 ContinuationBuilder.writeMemberType(VFPR); 2590 MemberCount++; 2591 continue; 2592 } 2593 2594 // Data member. 2595 uint64_t MemberOffsetInBits = 2596 Member->getOffsetInBits() + MemberInfo.BaseOffset; 2597 if (Member->isBitField()) { 2598 uint64_t StartBitOffset = MemberOffsetInBits; 2599 if (const auto *CI = 2600 dyn_cast_or_null<ConstantInt>(Member->getStorageOffsetInBits())) { 2601 MemberOffsetInBits = CI->getZExtValue() + MemberInfo.BaseOffset; 2602 } 2603 StartBitOffset -= MemberOffsetInBits; 2604 BitFieldRecord BFR(MemberBaseType, Member->getSizeInBits(), 2605 StartBitOffset); 2606 MemberBaseType = TypeTable.writeLeafType(BFR); 2607 } 2608 uint64_t MemberOffsetInBytes = MemberOffsetInBits / 8; 2609 DataMemberRecord DMR(Access, MemberBaseType, MemberOffsetInBytes, 2610 MemberName); 2611 ContinuationBuilder.writeMemberType(DMR); 2612 MemberCount++; 2613 } 2614 2615 // Create methods 2616 for (auto &MethodItr : Info.Methods) { 2617 StringRef Name = MethodItr.first->getString(); 2618 2619 std::vector<OneMethodRecord> Methods; 2620 for (const DISubprogram *SP : MethodItr.second) { 2621 TypeIndex MethodType = getMemberFunctionType(SP, Ty); 2622 bool Introduced = SP->getFlags() & DINode::FlagIntroducedVirtual; 2623 2624 unsigned VFTableOffset = -1; 2625 if (Introduced) 2626 VFTableOffset = SP->getVirtualIndex() * getPointerSizeInBytes(); 2627 2628 Methods.push_back(OneMethodRecord( 2629 MethodType, translateAccessFlags(Ty->getTag(), SP->getFlags()), 2630 translateMethodKindFlags(SP, Introduced), 2631 translateMethodOptionFlags(SP), VFTableOffset, Name)); 2632 MemberCount++; 2633 } 2634 assert(!Methods.empty() && "Empty methods map entry"); 2635 if (Methods.size() == 1) 2636 ContinuationBuilder.writeMemberType(Methods[0]); 2637 else { 2638 // FIXME: Make this use its own ContinuationBuilder so that 2639 // MethodOverloadList can be split correctly. 2640 MethodOverloadListRecord MOLR(Methods); 2641 TypeIndex MethodList = TypeTable.writeLeafType(MOLR); 2642 2643 OverloadedMethodRecord OMR(Methods.size(), MethodList, Name); 2644 ContinuationBuilder.writeMemberType(OMR); 2645 } 2646 } 2647 2648 // Create nested classes. 2649 for (const DIType *Nested : Info.NestedTypes) { 2650 NestedTypeRecord R(getTypeIndex(Nested), Nested->getName()); 2651 ContinuationBuilder.writeMemberType(R); 2652 MemberCount++; 2653 } 2654 2655 TypeIndex FieldTI = TypeTable.insertRecord(ContinuationBuilder); 2656 return std::make_tuple(FieldTI, Info.VShapeTI, MemberCount, 2657 !Info.NestedTypes.empty()); 2658 } 2659 2660 TypeIndex CodeViewDebug::getVBPTypeIndex() { 2661 if (!VBPType.getIndex()) { 2662 // Make a 'const int *' type. 2663 ModifierRecord MR(TypeIndex::Int32(), ModifierOptions::Const); 2664 TypeIndex ModifiedTI = TypeTable.writeLeafType(MR); 2665 2666 PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64 2667 : PointerKind::Near32; 2668 PointerMode PM = PointerMode::Pointer; 2669 PointerOptions PO = PointerOptions::None; 2670 PointerRecord PR(ModifiedTI, PK, PM, PO, getPointerSizeInBytes()); 2671 VBPType = TypeTable.writeLeafType(PR); 2672 } 2673 2674 return VBPType; 2675 } 2676 2677 TypeIndex CodeViewDebug::getTypeIndex(const DIType *Ty, const DIType *ClassTy) { 2678 // The null DIType is the void type. Don't try to hash it. 2679 if (!Ty) 2680 return TypeIndex::Void(); 2681 2682 // Check if we've already translated this type. Don't try to do a 2683 // get-or-create style insertion that caches the hash lookup across the 2684 // lowerType call. It will update the TypeIndices map. 2685 auto I = TypeIndices.find({Ty, ClassTy}); 2686 if (I != TypeIndices.end()) 2687 return I->second; 2688 2689 TypeLoweringScope S(*this); 2690 TypeIndex TI = lowerType(Ty, ClassTy); 2691 return recordTypeIndexForDINode(Ty, TI, ClassTy); 2692 } 2693 2694 codeview::TypeIndex 2695 CodeViewDebug::getTypeIndexForThisPtr(const DIDerivedType *PtrTy, 2696 const DISubroutineType *SubroutineTy) { 2697 assert(PtrTy->getTag() == dwarf::DW_TAG_pointer_type && 2698 "this type must be a pointer type"); 2699 2700 PointerOptions Options = PointerOptions::None; 2701 if (SubroutineTy->getFlags() & DINode::DIFlags::FlagLValueReference) 2702 Options = PointerOptions::LValueRefThisPointer; 2703 else if (SubroutineTy->getFlags() & DINode::DIFlags::FlagRValueReference) 2704 Options = PointerOptions::RValueRefThisPointer; 2705 2706 // Check if we've already translated this type. If there is no ref qualifier 2707 // on the function then we look up this pointer type with no associated class 2708 // so that the TypeIndex for the this pointer can be shared with the type 2709 // index for other pointers to this class type. If there is a ref qualifier 2710 // then we lookup the pointer using the subroutine as the parent type. 2711 auto I = TypeIndices.find({PtrTy, SubroutineTy}); 2712 if (I != TypeIndices.end()) 2713 return I->second; 2714 2715 TypeLoweringScope S(*this); 2716 TypeIndex TI = lowerTypePointer(PtrTy, Options); 2717 return recordTypeIndexForDINode(PtrTy, TI, SubroutineTy); 2718 } 2719 2720 TypeIndex CodeViewDebug::getTypeIndexForReferenceTo(const DIType *Ty) { 2721 PointerRecord PR(getTypeIndex(Ty), 2722 getPointerSizeInBytes() == 8 ? PointerKind::Near64 2723 : PointerKind::Near32, 2724 PointerMode::LValueReference, PointerOptions::None, 2725 Ty->getSizeInBits() / 8); 2726 return TypeTable.writeLeafType(PR); 2727 } 2728 2729 TypeIndex CodeViewDebug::getCompleteTypeIndex(const DIType *Ty) { 2730 // The null DIType is the void type. Don't try to hash it. 2731 if (!Ty) 2732 return TypeIndex::Void(); 2733 2734 // Look through typedefs when getting the complete type index. Call 2735 // getTypeIndex on the typdef to ensure that any UDTs are accumulated and are 2736 // emitted only once. 2737 if (Ty->getTag() == dwarf::DW_TAG_typedef) 2738 (void)getTypeIndex(Ty); 2739 while (Ty->getTag() == dwarf::DW_TAG_typedef) 2740 Ty = cast<DIDerivedType>(Ty)->getBaseType(); 2741 2742 // If this is a non-record type, the complete type index is the same as the 2743 // normal type index. Just call getTypeIndex. 2744 switch (Ty->getTag()) { 2745 case dwarf::DW_TAG_class_type: 2746 case dwarf::DW_TAG_structure_type: 2747 case dwarf::DW_TAG_union_type: 2748 break; 2749 default: 2750 return getTypeIndex(Ty); 2751 } 2752 2753 const auto *CTy = cast<DICompositeType>(Ty); 2754 2755 TypeLoweringScope S(*this); 2756 2757 // Make sure the forward declaration is emitted first. It's unclear if this 2758 // is necessary, but MSVC does it, and we should follow suit until we can show 2759 // otherwise. 2760 // We only emit a forward declaration for named types. 2761 if (!CTy->getName().empty() || !CTy->getIdentifier().empty()) { 2762 TypeIndex FwdDeclTI = getTypeIndex(CTy); 2763 2764 // Just use the forward decl if we don't have complete type info. This 2765 // might happen if the frontend is using modules and expects the complete 2766 // definition to be emitted elsewhere. 2767 if (CTy->isForwardDecl()) 2768 return FwdDeclTI; 2769 } 2770 2771 // Check if we've already translated the complete record type. 2772 // Insert the type with a null TypeIndex to signify that the type is currently 2773 // being lowered. 2774 auto InsertResult = CompleteTypeIndices.insert({CTy, TypeIndex()}); 2775 if (!InsertResult.second) 2776 return InsertResult.first->second; 2777 2778 TypeIndex TI; 2779 switch (CTy->getTag()) { 2780 case dwarf::DW_TAG_class_type: 2781 case dwarf::DW_TAG_structure_type: 2782 TI = lowerCompleteTypeClass(CTy); 2783 break; 2784 case dwarf::DW_TAG_union_type: 2785 TI = lowerCompleteTypeUnion(CTy); 2786 break; 2787 default: 2788 llvm_unreachable("not a record"); 2789 } 2790 2791 // Update the type index associated with this CompositeType. This cannot 2792 // use the 'InsertResult' iterator above because it is potentially 2793 // invalidated by map insertions which can occur while lowering the class 2794 // type above. 2795 CompleteTypeIndices[CTy] = TI; 2796 return TI; 2797 } 2798 2799 /// Emit all the deferred complete record types. Try to do this in FIFO order, 2800 /// and do this until fixpoint, as each complete record type typically 2801 /// references 2802 /// many other record types. 2803 void CodeViewDebug::emitDeferredCompleteTypes() { 2804 SmallVector<const DICompositeType *, 4> TypesToEmit; 2805 while (!DeferredCompleteTypes.empty()) { 2806 std::swap(DeferredCompleteTypes, TypesToEmit); 2807 for (const DICompositeType *RecordTy : TypesToEmit) 2808 getCompleteTypeIndex(RecordTy); 2809 TypesToEmit.clear(); 2810 } 2811 } 2812 2813 void CodeViewDebug::emitLocalVariableList(const FunctionInfo &FI, 2814 ArrayRef<LocalVariable> Locals) { 2815 // Get the sorted list of parameters and emit them first. 2816 SmallVector<const LocalVariable *, 6> Params; 2817 for (const LocalVariable &L : Locals) 2818 if (L.DIVar->isParameter()) 2819 Params.push_back(&L); 2820 llvm::sort(Params, [](const LocalVariable *L, const LocalVariable *R) { 2821 return L->DIVar->getArg() < R->DIVar->getArg(); 2822 }); 2823 for (const LocalVariable *L : Params) 2824 emitLocalVariable(FI, *L); 2825 2826 // Next emit all non-parameters in the order that we found them. 2827 for (const LocalVariable &L : Locals) { 2828 if (!L.DIVar->isParameter()) { 2829 if (L.ConstantValue) { 2830 // If ConstantValue is set we will emit it as a S_CONSTANT instead of a 2831 // S_LOCAL in order to be able to represent it at all. 2832 const DIType *Ty = L.DIVar->getType(); 2833 APSInt Val(*L.ConstantValue); 2834 emitConstantSymbolRecord(Ty, Val, std::string(L.DIVar->getName())); 2835 } else { 2836 emitLocalVariable(FI, L); 2837 } 2838 } 2839 } 2840 } 2841 2842 void CodeViewDebug::emitLocalVariable(const FunctionInfo &FI, 2843 const LocalVariable &Var) { 2844 // LocalSym record, see SymbolRecord.h for more info. 2845 MCSymbol *LocalEnd = beginSymbolRecord(SymbolKind::S_LOCAL); 2846 2847 LocalSymFlags Flags = LocalSymFlags::None; 2848 if (Var.DIVar->isParameter()) 2849 Flags |= LocalSymFlags::IsParameter; 2850 if (Var.DefRanges.empty()) 2851 Flags |= LocalSymFlags::IsOptimizedOut; 2852 2853 OS.AddComment("TypeIndex"); 2854 TypeIndex TI = Var.UseReferenceType 2855 ? getTypeIndexForReferenceTo(Var.DIVar->getType()) 2856 : getCompleteTypeIndex(Var.DIVar->getType()); 2857 OS.emitInt32(TI.getIndex()); 2858 OS.AddComment("Flags"); 2859 OS.emitInt16(static_cast<uint16_t>(Flags)); 2860 // Truncate the name so we won't overflow the record length field. 2861 emitNullTerminatedSymbolName(OS, Var.DIVar->getName()); 2862 endSymbolRecord(LocalEnd); 2863 2864 // Calculate the on disk prefix of the appropriate def range record. The 2865 // records and on disk formats are described in SymbolRecords.h. BytePrefix 2866 // should be big enough to hold all forms without memory allocation. 2867 SmallString<20> BytePrefix; 2868 for (const auto &Pair : Var.DefRanges) { 2869 LocalVarDef DefRange = Pair.first; 2870 const auto &Ranges = Pair.second; 2871 BytePrefix.clear(); 2872 if (DefRange.InMemory) { 2873 int Offset = DefRange.DataOffset; 2874 unsigned Reg = DefRange.CVRegister; 2875 2876 // 32-bit x86 call sequences often use PUSH instructions, which disrupt 2877 // ESP-relative offsets. Use the virtual frame pointer, VFRAME or $T0, 2878 // instead. In frames without stack realignment, $T0 will be the CFA. 2879 if (RegisterId(Reg) == RegisterId::ESP) { 2880 Reg = unsigned(RegisterId::VFRAME); 2881 Offset += FI.OffsetAdjustment; 2882 } 2883 2884 // If we can use the chosen frame pointer for the frame and this isn't a 2885 // sliced aggregate, use the smaller S_DEFRANGE_FRAMEPOINTER_REL record. 2886 // Otherwise, use S_DEFRANGE_REGISTER_REL. 2887 EncodedFramePtrReg EncFP = encodeFramePtrReg(RegisterId(Reg), TheCPU); 2888 if (!DefRange.IsSubfield && EncFP != EncodedFramePtrReg::None && 2889 (bool(Flags & LocalSymFlags::IsParameter) 2890 ? (EncFP == FI.EncodedParamFramePtrReg) 2891 : (EncFP == FI.EncodedLocalFramePtrReg))) { 2892 DefRangeFramePointerRelHeader DRHdr; 2893 DRHdr.Offset = Offset; 2894 OS.emitCVDefRangeDirective(Ranges, DRHdr); 2895 } else { 2896 uint16_t RegRelFlags = 0; 2897 if (DefRange.IsSubfield) { 2898 RegRelFlags = DefRangeRegisterRelSym::IsSubfieldFlag | 2899 (DefRange.StructOffset 2900 << DefRangeRegisterRelSym::OffsetInParentShift); 2901 } 2902 DefRangeRegisterRelHeader DRHdr; 2903 DRHdr.Register = Reg; 2904 DRHdr.Flags = RegRelFlags; 2905 DRHdr.BasePointerOffset = Offset; 2906 OS.emitCVDefRangeDirective(Ranges, DRHdr); 2907 } 2908 } else { 2909 assert(DefRange.DataOffset == 0 && "unexpected offset into register"); 2910 if (DefRange.IsSubfield) { 2911 DefRangeSubfieldRegisterHeader DRHdr; 2912 DRHdr.Register = DefRange.CVRegister; 2913 DRHdr.MayHaveNoName = 0; 2914 DRHdr.OffsetInParent = DefRange.StructOffset; 2915 OS.emitCVDefRangeDirective(Ranges, DRHdr); 2916 } else { 2917 DefRangeRegisterHeader DRHdr; 2918 DRHdr.Register = DefRange.CVRegister; 2919 DRHdr.MayHaveNoName = 0; 2920 OS.emitCVDefRangeDirective(Ranges, DRHdr); 2921 } 2922 } 2923 } 2924 } 2925 2926 void CodeViewDebug::emitLexicalBlockList(ArrayRef<LexicalBlock *> Blocks, 2927 const FunctionInfo& FI) { 2928 for (LexicalBlock *Block : Blocks) 2929 emitLexicalBlock(*Block, FI); 2930 } 2931 2932 /// Emit an S_BLOCK32 and S_END record pair delimiting the contents of a 2933 /// lexical block scope. 2934 void CodeViewDebug::emitLexicalBlock(const LexicalBlock &Block, 2935 const FunctionInfo& FI) { 2936 MCSymbol *RecordEnd = beginSymbolRecord(SymbolKind::S_BLOCK32); 2937 OS.AddComment("PtrParent"); 2938 OS.emitInt32(0); // PtrParent 2939 OS.AddComment("PtrEnd"); 2940 OS.emitInt32(0); // PtrEnd 2941 OS.AddComment("Code size"); 2942 OS.emitAbsoluteSymbolDiff(Block.End, Block.Begin, 4); // Code Size 2943 OS.AddComment("Function section relative address"); 2944 OS.emitCOFFSecRel32(Block.Begin, /*Offset=*/0); // Func Offset 2945 OS.AddComment("Function section index"); 2946 OS.emitCOFFSectionIndex(FI.Begin); // Func Symbol 2947 OS.AddComment("Lexical block name"); 2948 emitNullTerminatedSymbolName(OS, Block.Name); // Name 2949 endSymbolRecord(RecordEnd); 2950 2951 // Emit variables local to this lexical block. 2952 emitLocalVariableList(FI, Block.Locals); 2953 emitGlobalVariableList(Block.Globals); 2954 2955 // Emit lexical blocks contained within this block. 2956 emitLexicalBlockList(Block.Children, FI); 2957 2958 // Close the lexical block scope. 2959 emitEndSymbolRecord(SymbolKind::S_END); 2960 } 2961 2962 /// Convenience routine for collecting lexical block information for a list 2963 /// of lexical scopes. 2964 void CodeViewDebug::collectLexicalBlockInfo( 2965 SmallVectorImpl<LexicalScope *> &Scopes, 2966 SmallVectorImpl<LexicalBlock *> &Blocks, 2967 SmallVectorImpl<LocalVariable> &Locals, 2968 SmallVectorImpl<CVGlobalVariable> &Globals) { 2969 for (LexicalScope *Scope : Scopes) 2970 collectLexicalBlockInfo(*Scope, Blocks, Locals, Globals); 2971 } 2972 2973 /// Populate the lexical blocks and local variable lists of the parent with 2974 /// information about the specified lexical scope. 2975 void CodeViewDebug::collectLexicalBlockInfo( 2976 LexicalScope &Scope, 2977 SmallVectorImpl<LexicalBlock *> &ParentBlocks, 2978 SmallVectorImpl<LocalVariable> &ParentLocals, 2979 SmallVectorImpl<CVGlobalVariable> &ParentGlobals) { 2980 if (Scope.isAbstractScope()) 2981 return; 2982 2983 // Gather information about the lexical scope including local variables, 2984 // global variables, and address ranges. 2985 bool IgnoreScope = false; 2986 auto LI = ScopeVariables.find(&Scope); 2987 SmallVectorImpl<LocalVariable> *Locals = 2988 LI != ScopeVariables.end() ? &LI->second : nullptr; 2989 auto GI = ScopeGlobals.find(Scope.getScopeNode()); 2990 SmallVectorImpl<CVGlobalVariable> *Globals = 2991 GI != ScopeGlobals.end() ? GI->second.get() : nullptr; 2992 const DILexicalBlock *DILB = dyn_cast<DILexicalBlock>(Scope.getScopeNode()); 2993 const SmallVectorImpl<InsnRange> &Ranges = Scope.getRanges(); 2994 2995 // Ignore lexical scopes which do not contain variables. 2996 if (!Locals && !Globals) 2997 IgnoreScope = true; 2998 2999 // Ignore lexical scopes which are not lexical blocks. 3000 if (!DILB) 3001 IgnoreScope = true; 3002 3003 // Ignore scopes which have too many address ranges to represent in the 3004 // current CodeView format or do not have a valid address range. 3005 // 3006 // For lexical scopes with multiple address ranges you may be tempted to 3007 // construct a single range covering every instruction where the block is 3008 // live and everything in between. Unfortunately, Visual Studio only 3009 // displays variables from the first matching lexical block scope. If the 3010 // first lexical block contains exception handling code or cold code which 3011 // is moved to the bottom of the routine creating a single range covering 3012 // nearly the entire routine, then it will hide all other lexical blocks 3013 // and the variables they contain. 3014 if (Ranges.size() != 1 || !getLabelAfterInsn(Ranges.front().second)) 3015 IgnoreScope = true; 3016 3017 if (IgnoreScope) { 3018 // This scope can be safely ignored and eliminating it will reduce the 3019 // size of the debug information. Be sure to collect any variable and scope 3020 // information from the this scope or any of its children and collapse them 3021 // into the parent scope. 3022 if (Locals) 3023 ParentLocals.append(Locals->begin(), Locals->end()); 3024 if (Globals) 3025 ParentGlobals.append(Globals->begin(), Globals->end()); 3026 collectLexicalBlockInfo(Scope.getChildren(), 3027 ParentBlocks, 3028 ParentLocals, 3029 ParentGlobals); 3030 return; 3031 } 3032 3033 // Create a new CodeView lexical block for this lexical scope. If we've 3034 // seen this DILexicalBlock before then the scope tree is malformed and 3035 // we can handle this gracefully by not processing it a second time. 3036 auto BlockInsertion = CurFn->LexicalBlocks.insert({DILB, LexicalBlock()}); 3037 if (!BlockInsertion.second) 3038 return; 3039 3040 // Create a lexical block containing the variables and collect the 3041 // lexical block information for the children. 3042 const InsnRange &Range = Ranges.front(); 3043 assert(Range.first && Range.second); 3044 LexicalBlock &Block = BlockInsertion.first->second; 3045 Block.Begin = getLabelBeforeInsn(Range.first); 3046 Block.End = getLabelAfterInsn(Range.second); 3047 assert(Block.Begin && "missing label for scope begin"); 3048 assert(Block.End && "missing label for scope end"); 3049 Block.Name = DILB->getName(); 3050 if (Locals) 3051 Block.Locals = std::move(*Locals); 3052 if (Globals) 3053 Block.Globals = std::move(*Globals); 3054 ParentBlocks.push_back(&Block); 3055 collectLexicalBlockInfo(Scope.getChildren(), 3056 Block.Children, 3057 Block.Locals, 3058 Block.Globals); 3059 } 3060 3061 void CodeViewDebug::endFunctionImpl(const MachineFunction *MF) { 3062 const Function &GV = MF->getFunction(); 3063 assert(FnDebugInfo.count(&GV)); 3064 assert(CurFn == FnDebugInfo[&GV].get()); 3065 3066 collectVariableInfo(GV.getSubprogram()); 3067 3068 // Build the lexical block structure to emit for this routine. 3069 if (LexicalScope *CFS = LScopes.getCurrentFunctionScope()) 3070 collectLexicalBlockInfo(*CFS, 3071 CurFn->ChildBlocks, 3072 CurFn->Locals, 3073 CurFn->Globals); 3074 3075 // Clear the scope and variable information from the map which will not be 3076 // valid after we have finished processing this routine. This also prepares 3077 // the map for the subsequent routine. 3078 ScopeVariables.clear(); 3079 3080 // Don't emit anything if we don't have any line tables. 3081 // Thunks are compiler-generated and probably won't have source correlation. 3082 if (!CurFn->HaveLineInfo && !GV.getSubprogram()->isThunk()) { 3083 FnDebugInfo.erase(&GV); 3084 CurFn = nullptr; 3085 return; 3086 } 3087 3088 // Find heap alloc sites and add to list. 3089 for (const auto &MBB : *MF) { 3090 for (const auto &MI : MBB) { 3091 if (MDNode *MD = MI.getHeapAllocMarker()) { 3092 CurFn->HeapAllocSites.push_back(std::make_tuple(getLabelBeforeInsn(&MI), 3093 getLabelAfterInsn(&MI), 3094 dyn_cast<DIType>(MD))); 3095 } 3096 } 3097 } 3098 3099 bool isThumb = Triple(MMI->getModule()->getTargetTriple()).getArch() == 3100 llvm::Triple::ArchType::thumb; 3101 collectDebugInfoForJumpTables(MF, isThumb); 3102 3103 CurFn->Annotations = MF->getCodeViewAnnotations(); 3104 3105 CurFn->End = Asm->getFunctionEnd(); 3106 3107 CurFn = nullptr; 3108 } 3109 3110 // Usable locations are valid with non-zero line numbers. A line number of zero 3111 // corresponds to optimized code that doesn't have a distinct source location. 3112 // In this case, we try to use the previous or next source location depending on 3113 // the context. 3114 static bool isUsableDebugLoc(DebugLoc DL) { 3115 return DL && DL.getLine() != 0; 3116 } 3117 3118 void CodeViewDebug::beginInstruction(const MachineInstr *MI) { 3119 DebugHandlerBase::beginInstruction(MI); 3120 3121 // Ignore DBG_VALUE and DBG_LABEL locations and function prologue. 3122 if (!Asm || !CurFn || MI->isDebugInstr() || 3123 MI->getFlag(MachineInstr::FrameSetup)) 3124 return; 3125 3126 // If the first instruction of a new MBB has no location, find the first 3127 // instruction with a location and use that. 3128 DebugLoc DL = MI->getDebugLoc(); 3129 if (!isUsableDebugLoc(DL) && MI->getParent() != PrevInstBB) { 3130 for (const auto &NextMI : *MI->getParent()) { 3131 if (NextMI.isDebugInstr()) 3132 continue; 3133 DL = NextMI.getDebugLoc(); 3134 if (isUsableDebugLoc(DL)) 3135 break; 3136 } 3137 // FIXME: Handle the case where the BB has no valid locations. This would 3138 // probably require doing a real dataflow analysis. 3139 } 3140 PrevInstBB = MI->getParent(); 3141 3142 // If we still don't have a debug location, don't record a location. 3143 if (!isUsableDebugLoc(DL)) 3144 return; 3145 3146 maybeRecordLocation(DL, Asm->MF); 3147 } 3148 3149 MCSymbol *CodeViewDebug::beginCVSubsection(DebugSubsectionKind Kind) { 3150 MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(), 3151 *EndLabel = MMI->getContext().createTempSymbol(); 3152 OS.emitInt32(unsigned(Kind)); 3153 OS.AddComment("Subsection size"); 3154 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 4); 3155 OS.emitLabel(BeginLabel); 3156 return EndLabel; 3157 } 3158 3159 void CodeViewDebug::endCVSubsection(MCSymbol *EndLabel) { 3160 OS.emitLabel(EndLabel); 3161 // Every subsection must be aligned to a 4-byte boundary. 3162 OS.emitValueToAlignment(Align(4)); 3163 } 3164 3165 static StringRef getSymbolName(SymbolKind SymKind) { 3166 for (const EnumEntry<SymbolKind> &EE : getSymbolTypeNames()) 3167 if (EE.Value == SymKind) 3168 return EE.Name; 3169 return ""; 3170 } 3171 3172 MCSymbol *CodeViewDebug::beginSymbolRecord(SymbolKind SymKind) { 3173 MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(), 3174 *EndLabel = MMI->getContext().createTempSymbol(); 3175 OS.AddComment("Record length"); 3176 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2); 3177 OS.emitLabel(BeginLabel); 3178 if (OS.isVerboseAsm()) 3179 OS.AddComment("Record kind: " + getSymbolName(SymKind)); 3180 OS.emitInt16(unsigned(SymKind)); 3181 return EndLabel; 3182 } 3183 3184 void CodeViewDebug::endSymbolRecord(MCSymbol *SymEnd) { 3185 // MSVC does not pad out symbol records to four bytes, but LLVM does to avoid 3186 // an extra copy of every symbol record in LLD. This increases object file 3187 // size by less than 1% in the clang build, and is compatible with the Visual 3188 // C++ linker. 3189 OS.emitValueToAlignment(Align(4)); 3190 OS.emitLabel(SymEnd); 3191 } 3192 3193 void CodeViewDebug::emitEndSymbolRecord(SymbolKind EndKind) { 3194 OS.AddComment("Record length"); 3195 OS.emitInt16(2); 3196 if (OS.isVerboseAsm()) 3197 OS.AddComment("Record kind: " + getSymbolName(EndKind)); 3198 OS.emitInt16(uint16_t(EndKind)); // Record Kind 3199 } 3200 3201 void CodeViewDebug::emitDebugInfoForUDTs( 3202 const std::vector<std::pair<std::string, const DIType *>> &UDTs) { 3203 #ifndef NDEBUG 3204 size_t OriginalSize = UDTs.size(); 3205 #endif 3206 for (const auto &UDT : UDTs) { 3207 const DIType *T = UDT.second; 3208 assert(shouldEmitUdt(T)); 3209 MCSymbol *UDTRecordEnd = beginSymbolRecord(SymbolKind::S_UDT); 3210 OS.AddComment("Type"); 3211 OS.emitInt32(getCompleteTypeIndex(T).getIndex()); 3212 assert(OriginalSize == UDTs.size() && 3213 "getCompleteTypeIndex found new UDTs!"); 3214 emitNullTerminatedSymbolName(OS, UDT.first); 3215 endSymbolRecord(UDTRecordEnd); 3216 } 3217 } 3218 3219 void CodeViewDebug::collectGlobalVariableInfo() { 3220 DenseMap<const DIGlobalVariableExpression *, const GlobalVariable *> 3221 GlobalMap; 3222 for (const GlobalVariable &GV : MMI->getModule()->globals()) { 3223 SmallVector<DIGlobalVariableExpression *, 1> GVEs; 3224 GV.getDebugInfo(GVEs); 3225 for (const auto *GVE : GVEs) 3226 GlobalMap[GVE] = &GV; 3227 } 3228 3229 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 3230 for (const MDNode *Node : CUs->operands()) { 3231 const auto *CU = cast<DICompileUnit>(Node); 3232 for (const auto *GVE : CU->getGlobalVariables()) { 3233 const DIGlobalVariable *DIGV = GVE->getVariable(); 3234 const DIExpression *DIE = GVE->getExpression(); 3235 // Don't emit string literals in CodeView, as the only useful parts are 3236 // generally the filename and line number, which isn't possible to output 3237 // in CodeView. String literals should be the only unnamed GlobalVariable 3238 // with debug info. 3239 if (DIGV->getName().empty()) continue; 3240 3241 if ((DIE->getNumElements() == 2) && 3242 (DIE->getElement(0) == dwarf::DW_OP_plus_uconst)) 3243 // Record the constant offset for the variable. 3244 // 3245 // A Fortran common block uses this idiom to encode the offset 3246 // of a variable from the common block's starting address. 3247 CVGlobalVariableOffsets.insert( 3248 std::make_pair(DIGV, DIE->getElement(1))); 3249 3250 // Emit constant global variables in a global symbol section. 3251 if (GlobalMap.count(GVE) == 0 && DIE->isConstant()) { 3252 CVGlobalVariable CVGV = {DIGV, DIE}; 3253 GlobalVariables.emplace_back(std::move(CVGV)); 3254 } 3255 3256 const auto *GV = GlobalMap.lookup(GVE); 3257 if (!GV || GV->isDeclarationForLinker()) 3258 continue; 3259 3260 DIScope *Scope = DIGV->getScope(); 3261 SmallVector<CVGlobalVariable, 1> *VariableList; 3262 if (Scope && isa<DILocalScope>(Scope)) { 3263 // Locate a global variable list for this scope, creating one if 3264 // necessary. 3265 auto Insertion = ScopeGlobals.insert( 3266 {Scope, std::unique_ptr<GlobalVariableList>()}); 3267 if (Insertion.second) 3268 Insertion.first->second = std::make_unique<GlobalVariableList>(); 3269 VariableList = Insertion.first->second.get(); 3270 } else if (GV->hasComdat()) 3271 // Emit this global variable into a COMDAT section. 3272 VariableList = &ComdatVariables; 3273 else 3274 // Emit this global variable in a single global symbol section. 3275 VariableList = &GlobalVariables; 3276 CVGlobalVariable CVGV = {DIGV, GV}; 3277 VariableList->emplace_back(std::move(CVGV)); 3278 } 3279 } 3280 } 3281 3282 void CodeViewDebug::collectDebugInfoForGlobals() { 3283 for (const CVGlobalVariable &CVGV : GlobalVariables) { 3284 const DIGlobalVariable *DIGV = CVGV.DIGV; 3285 const DIScope *Scope = DIGV->getScope(); 3286 getCompleteTypeIndex(DIGV->getType()); 3287 getFullyQualifiedName(Scope, DIGV->getName()); 3288 } 3289 3290 for (const CVGlobalVariable &CVGV : ComdatVariables) { 3291 const DIGlobalVariable *DIGV = CVGV.DIGV; 3292 const DIScope *Scope = DIGV->getScope(); 3293 getCompleteTypeIndex(DIGV->getType()); 3294 getFullyQualifiedName(Scope, DIGV->getName()); 3295 } 3296 } 3297 3298 void CodeViewDebug::emitDebugInfoForGlobals() { 3299 // First, emit all globals that are not in a comdat in a single symbol 3300 // substream. MSVC doesn't like it if the substream is empty, so only open 3301 // it if we have at least one global to emit. 3302 switchToDebugSectionForSymbol(nullptr); 3303 if (!GlobalVariables.empty() || !StaticConstMembers.empty()) { 3304 OS.AddComment("Symbol subsection for globals"); 3305 MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols); 3306 emitGlobalVariableList(GlobalVariables); 3307 emitStaticConstMemberList(); 3308 endCVSubsection(EndLabel); 3309 } 3310 3311 // Second, emit each global that is in a comdat into its own .debug$S 3312 // section along with its own symbol substream. 3313 for (const CVGlobalVariable &CVGV : ComdatVariables) { 3314 const GlobalVariable *GV = cast<const GlobalVariable *>(CVGV.GVInfo); 3315 MCSymbol *GVSym = Asm->getSymbol(GV); 3316 OS.AddComment("Symbol subsection for " + 3317 Twine(GlobalValue::dropLLVMManglingEscape(GV->getName()))); 3318 switchToDebugSectionForSymbol(GVSym); 3319 MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols); 3320 // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. 3321 emitDebugInfoForGlobal(CVGV); 3322 endCVSubsection(EndLabel); 3323 } 3324 } 3325 3326 void CodeViewDebug::emitDebugInfoForRetainedTypes() { 3327 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 3328 for (const MDNode *Node : CUs->operands()) { 3329 for (auto *Ty : cast<DICompileUnit>(Node)->getRetainedTypes()) { 3330 if (DIType *RT = dyn_cast<DIType>(Ty)) { 3331 getTypeIndex(RT); 3332 // FIXME: Add to global/local DTU list. 3333 } 3334 } 3335 } 3336 } 3337 3338 // Emit each global variable in the specified array. 3339 void CodeViewDebug::emitGlobalVariableList(ArrayRef<CVGlobalVariable> Globals) { 3340 for (const CVGlobalVariable &CVGV : Globals) { 3341 // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. 3342 emitDebugInfoForGlobal(CVGV); 3343 } 3344 } 3345 3346 void CodeViewDebug::emitConstantSymbolRecord(const DIType *DTy, APSInt &Value, 3347 const std::string &QualifiedName) { 3348 MCSymbol *SConstantEnd = beginSymbolRecord(SymbolKind::S_CONSTANT); 3349 OS.AddComment("Type"); 3350 OS.emitInt32(getTypeIndex(DTy).getIndex()); 3351 3352 OS.AddComment("Value"); 3353 3354 // Encoded integers shouldn't need more than 10 bytes. 3355 uint8_t Data[10]; 3356 BinaryStreamWriter Writer(Data, llvm::support::endianness::little); 3357 CodeViewRecordIO IO(Writer); 3358 cantFail(IO.mapEncodedInteger(Value)); 3359 StringRef SRef((char *)Data, Writer.getOffset()); 3360 OS.emitBinaryData(SRef); 3361 3362 OS.AddComment("Name"); 3363 emitNullTerminatedSymbolName(OS, QualifiedName); 3364 endSymbolRecord(SConstantEnd); 3365 } 3366 3367 void CodeViewDebug::emitStaticConstMemberList() { 3368 for (const DIDerivedType *DTy : StaticConstMembers) { 3369 const DIScope *Scope = DTy->getScope(); 3370 3371 APSInt Value; 3372 if (const ConstantInt *CI = 3373 dyn_cast_or_null<ConstantInt>(DTy->getConstant())) 3374 Value = APSInt(CI->getValue(), 3375 DebugHandlerBase::isUnsignedDIType(DTy->getBaseType())); 3376 else if (const ConstantFP *CFP = 3377 dyn_cast_or_null<ConstantFP>(DTy->getConstant())) 3378 Value = APSInt(CFP->getValueAPF().bitcastToAPInt(), true); 3379 else 3380 llvm_unreachable("cannot emit a constant without a value"); 3381 3382 emitConstantSymbolRecord(DTy->getBaseType(), Value, 3383 getFullyQualifiedName(Scope, DTy->getName())); 3384 } 3385 } 3386 3387 static bool isFloatDIType(const DIType *Ty) { 3388 if (isa<DICompositeType>(Ty)) 3389 return false; 3390 3391 if (auto *DTy = dyn_cast<DIDerivedType>(Ty)) { 3392 dwarf::Tag T = (dwarf::Tag)Ty->getTag(); 3393 if (T == dwarf::DW_TAG_pointer_type || 3394 T == dwarf::DW_TAG_ptr_to_member_type || 3395 T == dwarf::DW_TAG_reference_type || 3396 T == dwarf::DW_TAG_rvalue_reference_type) 3397 return false; 3398 assert(DTy->getBaseType() && "Expected valid base type"); 3399 return isFloatDIType(DTy->getBaseType()); 3400 } 3401 3402 auto *BTy = cast<DIBasicType>(Ty); 3403 return (BTy->getEncoding() == dwarf::DW_ATE_float); 3404 } 3405 3406 void CodeViewDebug::emitDebugInfoForGlobal(const CVGlobalVariable &CVGV) { 3407 const DIGlobalVariable *DIGV = CVGV.DIGV; 3408 3409 const DIScope *Scope = DIGV->getScope(); 3410 // For static data members, get the scope from the declaration. 3411 if (const auto *MemberDecl = dyn_cast_or_null<DIDerivedType>( 3412 DIGV->getRawStaticDataMemberDeclaration())) 3413 Scope = MemberDecl->getScope(); 3414 // For static local variables and Fortran, the scoping portion is elided 3415 // in its name so that we can reference the variable in the command line 3416 // of the VS debugger. 3417 std::string QualifiedName = 3418 (moduleIsInFortran() || (Scope && isa<DILocalScope>(Scope))) 3419 ? std::string(DIGV->getName()) 3420 : getFullyQualifiedName(Scope, DIGV->getName()); 3421 3422 if (const GlobalVariable *GV = 3423 dyn_cast_if_present<const GlobalVariable *>(CVGV.GVInfo)) { 3424 // DataSym record, see SymbolRecord.h for more info. Thread local data 3425 // happens to have the same format as global data. 3426 MCSymbol *GVSym = Asm->getSymbol(GV); 3427 SymbolKind DataSym = GV->isThreadLocal() 3428 ? (DIGV->isLocalToUnit() ? SymbolKind::S_LTHREAD32 3429 : SymbolKind::S_GTHREAD32) 3430 : (DIGV->isLocalToUnit() ? SymbolKind::S_LDATA32 3431 : SymbolKind::S_GDATA32); 3432 MCSymbol *DataEnd = beginSymbolRecord(DataSym); 3433 OS.AddComment("Type"); 3434 OS.emitInt32(getCompleteTypeIndex(DIGV->getType()).getIndex()); 3435 OS.AddComment("DataOffset"); 3436 3437 uint64_t Offset = 0; 3438 if (CVGlobalVariableOffsets.contains(DIGV)) 3439 // Use the offset seen while collecting info on globals. 3440 Offset = CVGlobalVariableOffsets[DIGV]; 3441 OS.emitCOFFSecRel32(GVSym, Offset); 3442 3443 OS.AddComment("Segment"); 3444 OS.emitCOFFSectionIndex(GVSym); 3445 OS.AddComment("Name"); 3446 const unsigned LengthOfDataRecord = 12; 3447 emitNullTerminatedSymbolName(OS, QualifiedName, LengthOfDataRecord); 3448 endSymbolRecord(DataEnd); 3449 } else { 3450 const DIExpression *DIE = cast<const DIExpression *>(CVGV.GVInfo); 3451 assert(DIE->isConstant() && 3452 "Global constant variables must contain a constant expression."); 3453 3454 // Use unsigned for floats. 3455 bool isUnsigned = isFloatDIType(DIGV->getType()) 3456 ? true 3457 : DebugHandlerBase::isUnsignedDIType(DIGV->getType()); 3458 APSInt Value(APInt(/*BitWidth=*/64, DIE->getElement(1)), isUnsigned); 3459 emitConstantSymbolRecord(DIGV->getType(), Value, QualifiedName); 3460 } 3461 } 3462 3463 void forEachJumpTableBranch( 3464 const MachineFunction *MF, bool isThumb, 3465 const std::function<void(const MachineJumpTableInfo &, const MachineInstr &, 3466 int64_t)> &Callback) { 3467 auto JTI = MF->getJumpTableInfo(); 3468 if (JTI && !JTI->isEmpty()) { 3469 #ifndef NDEBUG 3470 auto UsedJTs = llvm::SmallBitVector(JTI->getJumpTables().size()); 3471 #endif 3472 for (const auto &MBB : *MF) { 3473 // Search for indirect branches... 3474 const auto LastMI = MBB.getFirstTerminator(); 3475 if (LastMI != MBB.end() && LastMI->isIndirectBranch()) { 3476 if (isThumb) { 3477 // ... that directly use jump table operands. 3478 // NOTE: ARM uses pattern matching to lower its BR_JT SDNode to 3479 // machine instructions, hence inserting a JUMP_TABLE_DEBUG_INFO node 3480 // interferes with this process *but* the resulting pseudo-instruction 3481 // uses a Jump Table operand, so extract the jump table index directly 3482 // from that. 3483 for (const auto &MO : LastMI->operands()) { 3484 if (MO.isJTI()) { 3485 unsigned Index = MO.getIndex(); 3486 #ifndef NDEBUG 3487 UsedJTs.set(Index); 3488 #endif 3489 Callback(*JTI, *LastMI, Index); 3490 break; 3491 } 3492 } 3493 } else { 3494 // ... that have jump table debug info. 3495 // NOTE: The debug info is inserted as a JUMP_TABLE_DEBUG_INFO node 3496 // when lowering the BR_JT SDNode to an indirect branch. 3497 for (auto I = MBB.instr_rbegin(), E = MBB.instr_rend(); I != E; ++I) { 3498 if (I->isJumpTableDebugInfo()) { 3499 unsigned Index = I->getOperand(0).getImm(); 3500 #ifndef NDEBUG 3501 UsedJTs.set(Index); 3502 #endif 3503 Callback(*JTI, *LastMI, Index); 3504 break; 3505 } 3506 } 3507 } 3508 } 3509 } 3510 #ifndef NDEBUG 3511 assert(UsedJTs.all() && 3512 "Some of jump tables were not used in a debug info instruction"); 3513 #endif 3514 } 3515 } 3516 3517 void CodeViewDebug::discoverJumpTableBranches(const MachineFunction *MF, 3518 bool isThumb) { 3519 forEachJumpTableBranch( 3520 MF, isThumb, 3521 [this](const MachineJumpTableInfo &, const MachineInstr &BranchMI, 3522 int64_t) { requestLabelBeforeInsn(&BranchMI); }); 3523 } 3524 3525 void CodeViewDebug::collectDebugInfoForJumpTables(const MachineFunction *MF, 3526 bool isThumb) { 3527 forEachJumpTableBranch( 3528 MF, isThumb, 3529 [this, MF](const MachineJumpTableInfo &JTI, const MachineInstr &BranchMI, 3530 int64_t JumpTableIndex) { 3531 // For label-difference jump tables, find the base expression. 3532 // Otherwise the jump table uses an absolute address (so no base 3533 // is required). 3534 const MCSymbol *Base; 3535 uint64_t BaseOffset = 0; 3536 const MCSymbol *Branch = getLabelBeforeInsn(&BranchMI); 3537 JumpTableEntrySize EntrySize; 3538 switch (JTI.getEntryKind()) { 3539 case MachineJumpTableInfo::EK_Custom32: 3540 case MachineJumpTableInfo::EK_GPRel32BlockAddress: 3541 case MachineJumpTableInfo::EK_GPRel64BlockAddress: 3542 llvm_unreachable( 3543 "EK_Custom32, EK_GPRel32BlockAddress, and " 3544 "EK_GPRel64BlockAddress should never be emitted for COFF"); 3545 case MachineJumpTableInfo::EK_BlockAddress: 3546 // Each entry is an absolute address. 3547 EntrySize = JumpTableEntrySize::Pointer; 3548 Base = nullptr; 3549 break; 3550 case MachineJumpTableInfo::EK_Inline: 3551 case MachineJumpTableInfo::EK_LabelDifference32: 3552 case MachineJumpTableInfo::EK_LabelDifference64: 3553 // Ask the AsmPrinter. 3554 std::tie(Base, BaseOffset, Branch, EntrySize) = 3555 Asm->getCodeViewJumpTableInfo(JumpTableIndex, &BranchMI, Branch); 3556 break; 3557 } 3558 3559 CurFn->JumpTables.push_back( 3560 {EntrySize, Base, BaseOffset, Branch, 3561 MF->getJTISymbol(JumpTableIndex, MMI->getContext()), 3562 JTI.getJumpTables()[JumpTableIndex].MBBs.size()}); 3563 }); 3564 } 3565 3566 void CodeViewDebug::emitDebugInfoForJumpTables(const FunctionInfo &FI) { 3567 for (auto JumpTable : FI.JumpTables) { 3568 MCSymbol *JumpTableEnd = beginSymbolRecord(SymbolKind::S_ARMSWITCHTABLE); 3569 if (JumpTable.Base) { 3570 OS.AddComment("Base offset"); 3571 OS.emitCOFFSecRel32(JumpTable.Base, JumpTable.BaseOffset); 3572 OS.AddComment("Base section index"); 3573 OS.emitCOFFSectionIndex(JumpTable.Base); 3574 } else { 3575 OS.AddComment("Base offset"); 3576 OS.emitInt32(0); 3577 OS.AddComment("Base section index"); 3578 OS.emitInt16(0); 3579 } 3580 OS.AddComment("Switch type"); 3581 OS.emitInt16(static_cast<uint16_t>(JumpTable.EntrySize)); 3582 OS.AddComment("Branch offset"); 3583 OS.emitCOFFSecRel32(JumpTable.Branch, /*Offset=*/0); 3584 OS.AddComment("Table offset"); 3585 OS.emitCOFFSecRel32(JumpTable.Table, /*Offset=*/0); 3586 OS.AddComment("Branch section index"); 3587 OS.emitCOFFSectionIndex(JumpTable.Branch); 3588 OS.AddComment("Table section index"); 3589 OS.emitCOFFSectionIndex(JumpTable.Table); 3590 OS.AddComment("Entries count"); 3591 OS.emitInt32(JumpTable.TableSize); 3592 endSymbolRecord(JumpTableEnd); 3593 } 3594 } 3595 3596 void CodeViewDebug::emitInlinees( 3597 const SmallSet<codeview::TypeIndex, 1> &Inlinees) { 3598 // Divide the list of inlinees into chunks such that each chunk fits within 3599 // one record. 3600 constexpr size_t ChunkSize = 3601 (MaxRecordLength - sizeof(SymbolKind) - sizeof(uint32_t)) / 3602 sizeof(uint32_t); 3603 3604 SmallVector<TypeIndex> SortedInlinees{Inlinees.begin(), Inlinees.end()}; 3605 llvm::sort(SortedInlinees); 3606 3607 size_t CurrentIndex = 0; 3608 while (CurrentIndex < SortedInlinees.size()) { 3609 auto Symbol = beginSymbolRecord(SymbolKind::S_INLINEES); 3610 auto CurrentChunkSize = 3611 std::min(ChunkSize, SortedInlinees.size() - CurrentIndex); 3612 OS.AddComment("Count"); 3613 OS.emitInt32(CurrentChunkSize); 3614 3615 const size_t CurrentChunkEnd = CurrentIndex + CurrentChunkSize; 3616 for (; CurrentIndex < CurrentChunkEnd; ++CurrentIndex) { 3617 OS.AddComment("Inlinee"); 3618 OS.emitInt32(SortedInlinees[CurrentIndex].getIndex()); 3619 } 3620 endSymbolRecord(Symbol); 3621 } 3622 } 3623