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