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