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/Program.h" 72 #include "llvm/Support/SMLoc.h" 73 #include "llvm/Support/ScopedPrinter.h" 74 #include "llvm/Target/TargetLoweringObjectFile.h" 75 #include "llvm/Target/TargetMachine.h" 76 #include <algorithm> 77 #include <cassert> 78 #include <cctype> 79 #include <cstddef> 80 #include <iterator> 81 #include <limits> 82 83 using namespace llvm; 84 using namespace llvm::codeview; 85 86 namespace { 87 class CVMCAdapter : public CodeViewRecordStreamer { 88 public: 89 CVMCAdapter(MCStreamer &OS, TypeCollection &TypeTable) 90 : OS(&OS), TypeTable(TypeTable) {} 91 92 void emitBytes(StringRef Data) override { OS->emitBytes(Data); } 93 94 void emitIntValue(uint64_t Value, unsigned Size) override { 95 OS->emitIntValueInHex(Value, Size); 96 } 97 98 void emitBinaryData(StringRef Data) override { OS->emitBinaryData(Data); } 99 100 void AddComment(const Twine &T) override { OS->AddComment(T); } 101 102 void AddRawComment(const Twine &T) override { OS->emitRawComment(T); } 103 104 bool isVerboseAsm() override { return OS->isVerboseAsm(); } 105 106 std::string getTypeName(TypeIndex TI) override { 107 std::string TypeName; 108 if (!TI.isNoneType()) { 109 if (TI.isSimple()) 110 TypeName = std::string(TypeIndex::simpleTypeName(TI)); 111 else 112 TypeName = std::string(TypeTable.getTypeName(TI)); 113 } 114 return TypeName; 115 } 116 117 private: 118 MCStreamer *OS = nullptr; 119 TypeCollection &TypeTable; 120 }; 121 } // namespace 122 123 static CPUType mapArchToCVCPUType(Triple::ArchType Type) { 124 switch (Type) { 125 case Triple::ArchType::x86: 126 return CPUType::Pentium3; 127 case Triple::ArchType::x86_64: 128 return CPUType::X64; 129 case Triple::ArchType::thumb: 130 return CPUType::Thumb; 131 case Triple::ArchType::aarch64: 132 return CPUType::ARM64; 133 default: 134 report_fatal_error("target architecture doesn't map to a CodeView CPUType"); 135 } 136 } 137 138 CodeViewDebug::CodeViewDebug(AsmPrinter *AP) 139 : DebugHandlerBase(AP), OS(*Asm->OutStreamer), TypeTable(Allocator) { 140 // If module doesn't have named metadata anchors or COFF debug section 141 // is not available, skip any debug info related stuff. 142 if (!MMI->getModule()->getNamedMetadata("llvm.dbg.cu") || 143 !AP->getObjFileLowering().getCOFFDebugSymbolsSection()) { 144 Asm = nullptr; 145 MMI->setDebugInfoAvailability(false); 146 return; 147 } 148 // Tell MMI that we have debug info. 149 MMI->setDebugInfoAvailability(true); 150 151 TheCPU = 152 mapArchToCVCPUType(Triple(MMI->getModule()->getTargetTriple()).getArch()); 153 154 collectGlobalVariableInfo(); 155 156 // Check if we should emit type record hashes. 157 ConstantInt *GH = mdconst::extract_or_null<ConstantInt>( 158 MMI->getModule()->getModuleFlag("CodeViewGHash")); 159 EmitDebugGlobalHashes = GH && !GH->isZero(); 160 } 161 162 StringRef CodeViewDebug::getFullFilepath(const DIFile *File) { 163 std::string &Filepath = FileToFilepathMap[File]; 164 if (!Filepath.empty()) 165 return Filepath; 166 167 StringRef Dir = File->getDirectory(), Filename = File->getFilename(); 168 169 // If this is a Unix-style path, just use it as is. Don't try to canonicalize 170 // it textually because one of the path components could be a symlink. 171 if (Dir.startswith("/") || Filename.startswith("/")) { 172 if (llvm::sys::path::is_absolute(Filename, llvm::sys::path::Style::posix)) 173 return Filename; 174 Filepath = std::string(Dir); 175 if (Dir.back() != '/') 176 Filepath += '/'; 177 Filepath += Filename; 178 return Filepath; 179 } 180 181 // Clang emits directory and relative filename info into the IR, but CodeView 182 // operates on full paths. We could change Clang to emit full paths too, but 183 // that would increase the IR size and probably not needed for other users. 184 // For now, just concatenate and canonicalize the path here. 185 if (Filename.find(':') == 1) 186 Filepath = std::string(Filename); 187 else 188 Filepath = (Dir + "\\" + Filename).str(); 189 190 // Canonicalize the path. We have to do it textually because we may no longer 191 // have access the file in the filesystem. 192 // First, replace all slashes with backslashes. 193 std::replace(Filepath.begin(), Filepath.end(), '/', '\\'); 194 195 // Remove all "\.\" with "\". 196 size_t Cursor = 0; 197 while ((Cursor = Filepath.find("\\.\\", Cursor)) != std::string::npos) 198 Filepath.erase(Cursor, 2); 199 200 // Replace all "\XXX\..\" with "\". Don't try too hard though as the original 201 // path should be well-formatted, e.g. start with a drive letter, etc. 202 Cursor = 0; 203 while ((Cursor = Filepath.find("\\..\\", Cursor)) != std::string::npos) { 204 // Something's wrong if the path starts with "\..\", abort. 205 if (Cursor == 0) 206 break; 207 208 size_t PrevSlash = Filepath.rfind('\\', Cursor - 1); 209 if (PrevSlash == std::string::npos) 210 // Something's wrong, abort. 211 break; 212 213 Filepath.erase(PrevSlash, Cursor + 3 - PrevSlash); 214 // The next ".." might be following the one we've just erased. 215 Cursor = PrevSlash; 216 } 217 218 // Remove all duplicate backslashes. 219 Cursor = 0; 220 while ((Cursor = Filepath.find("\\\\", Cursor)) != std::string::npos) 221 Filepath.erase(Cursor, 1); 222 223 return Filepath; 224 } 225 226 unsigned CodeViewDebug::maybeRecordFile(const DIFile *F) { 227 StringRef FullPath = getFullFilepath(F); 228 unsigned NextId = FileIdMap.size() + 1; 229 auto Insertion = FileIdMap.insert(std::make_pair(FullPath, NextId)); 230 if (Insertion.second) { 231 // We have to compute the full filepath and emit a .cv_file directive. 232 ArrayRef<uint8_t> ChecksumAsBytes; 233 FileChecksumKind CSKind = FileChecksumKind::None; 234 if (F->getChecksum()) { 235 std::string Checksum = fromHex(F->getChecksum()->Value); 236 void *CKMem = OS.getContext().allocate(Checksum.size(), 1); 237 memcpy(CKMem, Checksum.data(), Checksum.size()); 238 ChecksumAsBytes = ArrayRef<uint8_t>( 239 reinterpret_cast<const uint8_t *>(CKMem), Checksum.size()); 240 switch (F->getChecksum()->Kind) { 241 case DIFile::CSK_MD5: 242 CSKind = FileChecksumKind::MD5; 243 break; 244 case DIFile::CSK_SHA1: 245 CSKind = FileChecksumKind::SHA1; 246 break; 247 case DIFile::CSK_SHA256: 248 CSKind = FileChecksumKind::SHA256; 249 break; 250 } 251 } 252 bool Success = OS.EmitCVFileDirective(NextId, FullPath, ChecksumAsBytes, 253 static_cast<unsigned>(CSKind)); 254 (void)Success; 255 assert(Success && ".cv_file directive failed"); 256 } 257 return Insertion.first->second; 258 } 259 260 CodeViewDebug::InlineSite & 261 CodeViewDebug::getInlineSite(const DILocation *InlinedAt, 262 const DISubprogram *Inlinee) { 263 auto SiteInsertion = CurFn->InlineSites.insert({InlinedAt, InlineSite()}); 264 InlineSite *Site = &SiteInsertion.first->second; 265 if (SiteInsertion.second) { 266 unsigned ParentFuncId = CurFn->FuncId; 267 if (const DILocation *OuterIA = InlinedAt->getInlinedAt()) 268 ParentFuncId = 269 getInlineSite(OuterIA, InlinedAt->getScope()->getSubprogram()) 270 .SiteFuncId; 271 272 Site->SiteFuncId = NextFuncId++; 273 OS.EmitCVInlineSiteIdDirective( 274 Site->SiteFuncId, ParentFuncId, maybeRecordFile(InlinedAt->getFile()), 275 InlinedAt->getLine(), InlinedAt->getColumn(), SMLoc()); 276 Site->Inlinee = Inlinee; 277 InlinedSubprograms.insert(Inlinee); 278 getFuncIdForSubprogram(Inlinee); 279 } 280 return *Site; 281 } 282 283 static StringRef getPrettyScopeName(const DIScope *Scope) { 284 StringRef ScopeName = Scope->getName(); 285 if (!ScopeName.empty()) 286 return ScopeName; 287 288 switch (Scope->getTag()) { 289 case dwarf::DW_TAG_enumeration_type: 290 case dwarf::DW_TAG_class_type: 291 case dwarf::DW_TAG_structure_type: 292 case dwarf::DW_TAG_union_type: 293 return "<unnamed-tag>"; 294 case dwarf::DW_TAG_namespace: 295 return "`anonymous namespace'"; 296 } 297 298 return StringRef(); 299 } 300 301 const DISubprogram *CodeViewDebug::collectParentScopeNames( 302 const DIScope *Scope, SmallVectorImpl<StringRef> &QualifiedNameComponents) { 303 const DISubprogram *ClosestSubprogram = nullptr; 304 while (Scope != nullptr) { 305 if (ClosestSubprogram == nullptr) 306 ClosestSubprogram = dyn_cast<DISubprogram>(Scope); 307 308 // If a type appears in a scope chain, make sure it gets emitted. The 309 // frontend will be responsible for deciding if this should be a forward 310 // declaration or a complete type. 311 if (const auto *Ty = dyn_cast<DICompositeType>(Scope)) 312 DeferredCompleteTypes.push_back(Ty); 313 314 StringRef ScopeName = getPrettyScopeName(Scope); 315 if (!ScopeName.empty()) 316 QualifiedNameComponents.push_back(ScopeName); 317 Scope = Scope->getScope(); 318 } 319 return ClosestSubprogram; 320 } 321 322 static std::string formatNestedName(ArrayRef<StringRef> QualifiedNameComponents, 323 StringRef TypeName) { 324 std::string FullyQualifiedName; 325 for (StringRef QualifiedNameComponent : 326 llvm::reverse(QualifiedNameComponents)) { 327 FullyQualifiedName.append(std::string(QualifiedNameComponent)); 328 FullyQualifiedName.append("::"); 329 } 330 FullyQualifiedName.append(std::string(TypeName)); 331 return FullyQualifiedName; 332 } 333 334 struct CodeViewDebug::TypeLoweringScope { 335 TypeLoweringScope(CodeViewDebug &CVD) : CVD(CVD) { ++CVD.TypeEmissionLevel; } 336 ~TypeLoweringScope() { 337 // Don't decrement TypeEmissionLevel until after emitting deferred types, so 338 // inner TypeLoweringScopes don't attempt to emit deferred types. 339 if (CVD.TypeEmissionLevel == 1) 340 CVD.emitDeferredCompleteTypes(); 341 --CVD.TypeEmissionLevel; 342 } 343 CodeViewDebug &CVD; 344 }; 345 346 std::string CodeViewDebug::getFullyQualifiedName(const DIScope *Scope, 347 StringRef Name) { 348 // Ensure types in the scope chain are emitted as soon as possible. 349 // This can create otherwise a situation where S_UDTs are emitted while 350 // looping in emitDebugInfoForUDTs. 351 TypeLoweringScope S(*this); 352 SmallVector<StringRef, 5> QualifiedNameComponents; 353 collectParentScopeNames(Scope, QualifiedNameComponents); 354 return formatNestedName(QualifiedNameComponents, Name); 355 } 356 357 std::string CodeViewDebug::getFullyQualifiedName(const DIScope *Ty) { 358 const DIScope *Scope = Ty->getScope(); 359 return getFullyQualifiedName(Scope, getPrettyScopeName(Ty)); 360 } 361 362 TypeIndex CodeViewDebug::getScopeIndex(const DIScope *Scope) { 363 // No scope means global scope and that uses the zero index. 364 if (!Scope || isa<DIFile>(Scope)) 365 return TypeIndex(); 366 367 assert(!isa<DIType>(Scope) && "shouldn't make a namespace scope for a type"); 368 369 // Check if we've already translated this scope. 370 auto I = TypeIndices.find({Scope, nullptr}); 371 if (I != TypeIndices.end()) 372 return I->second; 373 374 // Build the fully qualified name of the scope. 375 std::string ScopeName = getFullyQualifiedName(Scope); 376 StringIdRecord SID(TypeIndex(), ScopeName); 377 auto TI = TypeTable.writeLeafType(SID); 378 return recordTypeIndexForDINode(Scope, TI); 379 } 380 381 TypeIndex CodeViewDebug::getFuncIdForSubprogram(const DISubprogram *SP) { 382 assert(SP); 383 384 // Check if we've already translated this subprogram. 385 auto I = TypeIndices.find({SP, nullptr}); 386 if (I != TypeIndices.end()) 387 return I->second; 388 389 // The display name includes function template arguments. Drop them to match 390 // MSVC. 391 StringRef DisplayName = SP->getName().split('<').first; 392 393 const DIScope *Scope = SP->getScope(); 394 TypeIndex TI; 395 if (const auto *Class = dyn_cast_or_null<DICompositeType>(Scope)) { 396 // If the scope is a DICompositeType, then this must be a method. Member 397 // function types take some special handling, and require access to the 398 // subprogram. 399 TypeIndex ClassType = getTypeIndex(Class); 400 MemberFuncIdRecord MFuncId(ClassType, getMemberFunctionType(SP, Class), 401 DisplayName); 402 TI = TypeTable.writeLeafType(MFuncId); 403 } else { 404 // Otherwise, this must be a free function. 405 TypeIndex ParentScope = getScopeIndex(Scope); 406 FuncIdRecord FuncId(ParentScope, getTypeIndex(SP->getType()), DisplayName); 407 TI = TypeTable.writeLeafType(FuncId); 408 } 409 410 return recordTypeIndexForDINode(SP, TI); 411 } 412 413 static bool isNonTrivial(const DICompositeType *DCTy) { 414 return ((DCTy->getFlags() & DINode::FlagNonTrivial) == DINode::FlagNonTrivial); 415 } 416 417 static FunctionOptions 418 getFunctionOptions(const DISubroutineType *Ty, 419 const DICompositeType *ClassTy = nullptr, 420 StringRef SPName = StringRef("")) { 421 FunctionOptions FO = FunctionOptions::None; 422 const DIType *ReturnTy = nullptr; 423 if (auto TypeArray = Ty->getTypeArray()) { 424 if (TypeArray.size()) 425 ReturnTy = TypeArray[0]; 426 } 427 428 // Add CxxReturnUdt option to functions that return nontrivial record types 429 // or methods that return record types. 430 if (auto *ReturnDCTy = dyn_cast_or_null<DICompositeType>(ReturnTy)) 431 if (isNonTrivial(ReturnDCTy) || ClassTy) 432 FO |= FunctionOptions::CxxReturnUdt; 433 434 // DISubroutineType is unnamed. Use DISubprogram's i.e. SPName in comparison. 435 if (ClassTy && isNonTrivial(ClassTy) && SPName == ClassTy->getName()) { 436 FO |= FunctionOptions::Constructor; 437 438 // TODO: put the FunctionOptions::ConstructorWithVirtualBases flag. 439 440 } 441 return FO; 442 } 443 444 TypeIndex CodeViewDebug::getMemberFunctionType(const DISubprogram *SP, 445 const DICompositeType *Class) { 446 // Always use the method declaration as the key for the function type. The 447 // method declaration contains the this adjustment. 448 if (SP->getDeclaration()) 449 SP = SP->getDeclaration(); 450 assert(!SP->getDeclaration() && "should use declaration as key"); 451 452 // Key the MemberFunctionRecord into the map as {SP, Class}. It won't collide 453 // with the MemberFuncIdRecord, which is keyed in as {SP, nullptr}. 454 auto I = TypeIndices.find({SP, Class}); 455 if (I != TypeIndices.end()) 456 return I->second; 457 458 // Make sure complete type info for the class is emitted *after* the member 459 // function type, as the complete class type is likely to reference this 460 // member function type. 461 TypeLoweringScope S(*this); 462 const bool IsStaticMethod = (SP->getFlags() & DINode::FlagStaticMember) != 0; 463 464 FunctionOptions FO = getFunctionOptions(SP->getType(), Class, SP->getName()); 465 TypeIndex TI = lowerTypeMemberFunction( 466 SP->getType(), Class, SP->getThisAdjustment(), IsStaticMethod, FO); 467 return recordTypeIndexForDINode(SP, TI, Class); 468 } 469 470 TypeIndex CodeViewDebug::recordTypeIndexForDINode(const DINode *Node, 471 TypeIndex TI, 472 const DIType *ClassTy) { 473 auto InsertResult = TypeIndices.insert({{Node, ClassTy}, TI}); 474 (void)InsertResult; 475 assert(InsertResult.second && "DINode was already assigned a type index"); 476 return TI; 477 } 478 479 unsigned CodeViewDebug::getPointerSizeInBytes() { 480 return MMI->getModule()->getDataLayout().getPointerSizeInBits() / 8; 481 } 482 483 void CodeViewDebug::recordLocalVariable(LocalVariable &&Var, 484 const LexicalScope *LS) { 485 if (const DILocation *InlinedAt = LS->getInlinedAt()) { 486 // This variable was inlined. Associate it with the InlineSite. 487 const DISubprogram *Inlinee = Var.DIVar->getScope()->getSubprogram(); 488 InlineSite &Site = getInlineSite(InlinedAt, Inlinee); 489 Site.InlinedLocals.emplace_back(Var); 490 } else { 491 // This variable goes into the corresponding lexical scope. 492 ScopeVariables[LS].emplace_back(Var); 493 } 494 } 495 496 static void addLocIfNotPresent(SmallVectorImpl<const DILocation *> &Locs, 497 const DILocation *Loc) { 498 if (!llvm::is_contained(Locs, Loc)) 499 Locs.push_back(Loc); 500 } 501 502 void CodeViewDebug::maybeRecordLocation(const DebugLoc &DL, 503 const MachineFunction *MF) { 504 // Skip this instruction if it has the same location as the previous one. 505 if (!DL || DL == PrevInstLoc) 506 return; 507 508 const DIScope *Scope = DL.get()->getScope(); 509 if (!Scope) 510 return; 511 512 // Skip this line if it is longer than the maximum we can record. 513 LineInfo LI(DL.getLine(), DL.getLine(), /*IsStatement=*/true); 514 if (LI.getStartLine() != DL.getLine() || LI.isAlwaysStepInto() || 515 LI.isNeverStepInto()) 516 return; 517 518 ColumnInfo CI(DL.getCol(), /*EndColumn=*/0); 519 if (CI.getStartColumn() != DL.getCol()) 520 return; 521 522 if (!CurFn->HaveLineInfo) 523 CurFn->HaveLineInfo = true; 524 unsigned FileId = 0; 525 if (PrevInstLoc.get() && PrevInstLoc->getFile() == DL->getFile()) 526 FileId = CurFn->LastFileId; 527 else 528 FileId = CurFn->LastFileId = maybeRecordFile(DL->getFile()); 529 PrevInstLoc = DL; 530 531 unsigned FuncId = CurFn->FuncId; 532 if (const DILocation *SiteLoc = DL->getInlinedAt()) { 533 const DILocation *Loc = DL.get(); 534 535 // If this location was actually inlined from somewhere else, give it the ID 536 // of the inline call site. 537 FuncId = 538 getInlineSite(SiteLoc, Loc->getScope()->getSubprogram()).SiteFuncId; 539 540 // Ensure we have links in the tree of inline call sites. 541 bool FirstLoc = true; 542 while ((SiteLoc = Loc->getInlinedAt())) { 543 InlineSite &Site = 544 getInlineSite(SiteLoc, Loc->getScope()->getSubprogram()); 545 if (!FirstLoc) 546 addLocIfNotPresent(Site.ChildSites, Loc); 547 FirstLoc = false; 548 Loc = SiteLoc; 549 } 550 addLocIfNotPresent(CurFn->ChildSites, Loc); 551 } 552 553 OS.emitCVLocDirective(FuncId, FileId, DL.getLine(), DL.getCol(), 554 /*PrologueEnd=*/false, /*IsStmt=*/false, 555 DL->getFilename(), SMLoc()); 556 } 557 558 void CodeViewDebug::emitCodeViewMagicVersion() { 559 OS.emitValueToAlignment(4); 560 OS.AddComment("Debug section magic"); 561 OS.emitInt32(COFF::DEBUG_SECTION_MAGIC); 562 } 563 564 void CodeViewDebug::endModule() { 565 if (!Asm || !MMI->hasDebugInfo()) 566 return; 567 568 assert(Asm != nullptr); 569 570 // The COFF .debug$S section consists of several subsections, each starting 571 // with a 4-byte control code (e.g. 0xF1, 0xF2, etc) and then a 4-byte length 572 // of the payload followed by the payload itself. The subsections are 4-byte 573 // aligned. 574 575 // Use the generic .debug$S section, and make a subsection for all the inlined 576 // subprograms. 577 switchToDebugSectionForSymbol(nullptr); 578 579 MCSymbol *CompilerInfo = beginCVSubsection(DebugSubsectionKind::Symbols); 580 emitCompilerInformation(); 581 endCVSubsection(CompilerInfo); 582 583 emitInlineeLinesSubsection(); 584 585 // Emit per-function debug information. 586 for (auto &P : FnDebugInfo) 587 if (!P.first->isDeclarationForLinker()) 588 emitDebugInfoForFunction(P.first, *P.second); 589 590 // Emit global variable debug information. 591 setCurrentSubprogram(nullptr); 592 emitDebugInfoForGlobals(); 593 594 // Emit retained types. 595 emitDebugInfoForRetainedTypes(); 596 597 // Switch back to the generic .debug$S section after potentially processing 598 // comdat symbol sections. 599 switchToDebugSectionForSymbol(nullptr); 600 601 // Emit UDT records for any types used by global variables. 602 if (!GlobalUDTs.empty()) { 603 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 604 emitDebugInfoForUDTs(GlobalUDTs); 605 endCVSubsection(SymbolsEnd); 606 } 607 608 // This subsection holds a file index to offset in string table table. 609 OS.AddComment("File index to string table offset subsection"); 610 OS.emitCVFileChecksumsDirective(); 611 612 // This subsection holds the string table. 613 OS.AddComment("String table"); 614 OS.emitCVStringTableDirective(); 615 616 // Emit S_BUILDINFO, which points to LF_BUILDINFO. Put this in its own symbol 617 // subsection in the generic .debug$S section at the end. There is no 618 // particular reason for this ordering other than to match MSVC. 619 emitBuildInfo(); 620 621 // Emit type information and hashes last, so that any types we translate while 622 // emitting function info are included. 623 emitTypeInformation(); 624 625 if (EmitDebugGlobalHashes) 626 emitTypeGlobalHashes(); 627 628 clear(); 629 } 630 631 static void 632 emitNullTerminatedSymbolName(MCStreamer &OS, StringRef S, 633 unsigned MaxFixedRecordLength = 0xF00) { 634 // The maximum CV record length is 0xFF00. Most of the strings we emit appear 635 // after a fixed length portion of the record. The fixed length portion should 636 // always be less than 0xF00 (3840) bytes, so truncate the string so that the 637 // overall record size is less than the maximum allowed. 638 SmallString<32> NullTerminatedString( 639 S.take_front(MaxRecordLength - MaxFixedRecordLength - 1)); 640 NullTerminatedString.push_back('\0'); 641 OS.emitBytes(NullTerminatedString); 642 } 643 644 void CodeViewDebug::emitTypeInformation() { 645 if (TypeTable.empty()) 646 return; 647 648 // Start the .debug$T or .debug$P section with 0x4. 649 OS.SwitchSection(Asm->getObjFileLowering().getCOFFDebugTypesSection()); 650 emitCodeViewMagicVersion(); 651 652 TypeTableCollection Table(TypeTable.records()); 653 TypeVisitorCallbackPipeline Pipeline; 654 655 // To emit type record using Codeview MCStreamer adapter 656 CVMCAdapter CVMCOS(OS, Table); 657 TypeRecordMapping typeMapping(CVMCOS); 658 Pipeline.addCallbackToPipeline(typeMapping); 659 660 Optional<TypeIndex> B = Table.getFirst(); 661 while (B) { 662 // This will fail if the record data is invalid. 663 CVType Record = Table.getType(*B); 664 665 Error E = codeview::visitTypeRecord(Record, *B, Pipeline); 666 667 if (E) { 668 logAllUnhandledErrors(std::move(E), errs(), "error: "); 669 llvm_unreachable("produced malformed type record"); 670 } 671 672 B = Table.getNext(*B); 673 } 674 } 675 676 void CodeViewDebug::emitTypeGlobalHashes() { 677 if (TypeTable.empty()) 678 return; 679 680 // Start the .debug$H section with the version and hash algorithm, currently 681 // hardcoded to version 0, SHA1. 682 OS.SwitchSection(Asm->getObjFileLowering().getCOFFGlobalTypeHashesSection()); 683 684 OS.emitValueToAlignment(4); 685 OS.AddComment("Magic"); 686 OS.emitInt32(COFF::DEBUG_HASHES_SECTION_MAGIC); 687 OS.AddComment("Section Version"); 688 OS.emitInt16(0); 689 OS.AddComment("Hash Algorithm"); 690 OS.emitInt16(uint16_t(GlobalTypeHashAlg::SHA1_8)); 691 692 TypeIndex TI(TypeIndex::FirstNonSimpleIndex); 693 for (const auto &GHR : TypeTable.hashes()) { 694 if (OS.isVerboseAsm()) { 695 // Emit an EOL-comment describing which TypeIndex this hash corresponds 696 // to, as well as the stringified SHA1 hash. 697 SmallString<32> Comment; 698 raw_svector_ostream CommentOS(Comment); 699 CommentOS << formatv("{0:X+} [{1}]", TI.getIndex(), GHR); 700 OS.AddComment(Comment); 701 ++TI; 702 } 703 assert(GHR.Hash.size() == 8); 704 StringRef S(reinterpret_cast<const char *>(GHR.Hash.data()), 705 GHR.Hash.size()); 706 OS.emitBinaryData(S); 707 } 708 } 709 710 static SourceLanguage MapDWLangToCVLang(unsigned DWLang) { 711 switch (DWLang) { 712 case dwarf::DW_LANG_C: 713 case dwarf::DW_LANG_C89: 714 case dwarf::DW_LANG_C99: 715 case dwarf::DW_LANG_C11: 716 case dwarf::DW_LANG_ObjC: 717 return SourceLanguage::C; 718 case dwarf::DW_LANG_C_plus_plus: 719 case dwarf::DW_LANG_C_plus_plus_03: 720 case dwarf::DW_LANG_C_plus_plus_11: 721 case dwarf::DW_LANG_C_plus_plus_14: 722 return SourceLanguage::Cpp; 723 case dwarf::DW_LANG_Fortran77: 724 case dwarf::DW_LANG_Fortran90: 725 case dwarf::DW_LANG_Fortran03: 726 case dwarf::DW_LANG_Fortran08: 727 return SourceLanguage::Fortran; 728 case dwarf::DW_LANG_Pascal83: 729 return SourceLanguage::Pascal; 730 case dwarf::DW_LANG_Cobol74: 731 case dwarf::DW_LANG_Cobol85: 732 return SourceLanguage::Cobol; 733 case dwarf::DW_LANG_Java: 734 return SourceLanguage::Java; 735 case dwarf::DW_LANG_D: 736 return SourceLanguage::D; 737 case dwarf::DW_LANG_Swift: 738 return SourceLanguage::Swift; 739 default: 740 // There's no CodeView representation for this language, and CV doesn't 741 // have an "unknown" option for the language field, so we'll use MASM, 742 // as it's very low level. 743 return SourceLanguage::Masm; 744 } 745 } 746 747 namespace { 748 struct Version { 749 int Part[4]; 750 }; 751 } // end anonymous namespace 752 753 // Takes a StringRef like "clang 4.0.0.0 (other nonsense 123)" and parses out 754 // the version number. 755 static Version parseVersion(StringRef Name) { 756 Version V = {{0}}; 757 int N = 0; 758 for (const char C : Name) { 759 if (isdigit(C)) { 760 V.Part[N] *= 10; 761 V.Part[N] += C - '0'; 762 } else if (C == '.') { 763 ++N; 764 if (N >= 4) 765 return V; 766 } else if (N > 0) 767 return V; 768 } 769 return V; 770 } 771 772 void CodeViewDebug::emitCompilerInformation() { 773 MCSymbol *CompilerEnd = beginSymbolRecord(SymbolKind::S_COMPILE3); 774 uint32_t Flags = 0; 775 776 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 777 const MDNode *Node = *CUs->operands().begin(); 778 const auto *CU = cast<DICompileUnit>(Node); 779 780 // The low byte of the flags indicates the source language. 781 Flags = MapDWLangToCVLang(CU->getSourceLanguage()); 782 // TODO: Figure out which other flags need to be set. 783 784 OS.AddComment("Flags and language"); 785 OS.emitInt32(Flags); 786 787 OS.AddComment("CPUType"); 788 OS.emitInt16(static_cast<uint64_t>(TheCPU)); 789 790 StringRef CompilerVersion = CU->getProducer(); 791 Version FrontVer = parseVersion(CompilerVersion); 792 OS.AddComment("Frontend version"); 793 for (int N = 0; N < 4; ++N) 794 OS.emitInt16(FrontVer.Part[N]); 795 796 // Some Microsoft tools, like Binscope, expect a backend version number of at 797 // least 8.something, so we'll coerce the LLVM version into a form that 798 // guarantees it'll be big enough without really lying about the version. 799 int Major = 1000 * LLVM_VERSION_MAJOR + 800 10 * LLVM_VERSION_MINOR + 801 LLVM_VERSION_PATCH; 802 // Clamp it for builds that use unusually large version numbers. 803 Major = std::min<int>(Major, std::numeric_limits<uint16_t>::max()); 804 Version BackVer = {{ Major, 0, 0, 0 }}; 805 OS.AddComment("Backend version"); 806 for (int N = 0; N < 4; ++N) 807 OS.emitInt16(BackVer.Part[N]); 808 809 OS.AddComment("Null-terminated compiler version string"); 810 emitNullTerminatedSymbolName(OS, CompilerVersion); 811 812 endSymbolRecord(CompilerEnd); 813 } 814 815 static TypeIndex getStringIdTypeIdx(GlobalTypeTableBuilder &TypeTable, 816 StringRef S) { 817 StringIdRecord SIR(TypeIndex(0x0), S); 818 return TypeTable.writeLeafType(SIR); 819 } 820 821 static std::string flattenCommandLine(ArrayRef<const char *> Args, 822 StringRef MainFilename) { 823 std::string FlatCmdLine; 824 raw_string_ostream OS(FlatCmdLine); 825 StringRef LastArg; 826 for (StringRef Arg : Args) { 827 if (Arg.empty()) 828 continue; 829 // The command-line shall not contain the file to compile. 830 if (Arg == MainFilename && LastArg != "-main-file-name") 831 continue; 832 // Also remove the output file. 833 if (Arg == "-o" || LastArg == "-o") { 834 LastArg = Arg; 835 continue; 836 } 837 if (!LastArg.empty()) 838 OS << " "; 839 llvm::sys::printArg(OS, Arg, /*Quote=*/true); 840 LastArg = Arg; 841 } 842 OS.flush(); 843 return FlatCmdLine; 844 } 845 846 void CodeViewDebug::emitBuildInfo() { 847 // First, make LF_BUILDINFO. It's a sequence of strings with various bits of 848 // build info. The known prefix is: 849 // - Absolute path of current directory 850 // - Compiler path 851 // - Main source file path, relative to CWD or absolute 852 // - Type server PDB file 853 // - Canonical compiler command line 854 // If frontend and backend compilation are separated (think llc or LTO), it's 855 // not clear if the compiler path should refer to the executable for the 856 // frontend or the backend. Leave it blank for now. 857 TypeIndex BuildInfoArgs[BuildInfoRecord::MaxArgs] = {}; 858 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 859 const MDNode *Node = *CUs->operands().begin(); // FIXME: Multiple CUs. 860 const auto *CU = cast<DICompileUnit>(Node); 861 const DIFile *MainSourceFile = CU->getFile(); 862 BuildInfoArgs[BuildInfoRecord::CurrentDirectory] = 863 getStringIdTypeIdx(TypeTable, MainSourceFile->getDirectory()); 864 BuildInfoArgs[BuildInfoRecord::SourceFile] = 865 getStringIdTypeIdx(TypeTable, MainSourceFile->getFilename()); 866 // FIXME: PDB is intentionally blank unless we implement /Zi type servers. 867 BuildInfoArgs[BuildInfoRecord::TypeServerPDB] = 868 getStringIdTypeIdx(TypeTable, ""); 869 if (Asm->TM.Options.MCOptions.Argv0 != nullptr) { 870 BuildInfoArgs[BuildInfoRecord::BuildTool] = 871 getStringIdTypeIdx(TypeTable, Asm->TM.Options.MCOptions.Argv0); 872 BuildInfoArgs[BuildInfoRecord::CommandLine] = getStringIdTypeIdx( 873 TypeTable, flattenCommandLine(Asm->TM.Options.MCOptions.CommandLineArgs, 874 MainSourceFile->getFilename())); 875 } 876 BuildInfoRecord BIR(BuildInfoArgs); 877 TypeIndex BuildInfoIndex = TypeTable.writeLeafType(BIR); 878 879 // Make a new .debug$S subsection for the S_BUILDINFO record, which points 880 // from the module symbols into the type stream. 881 MCSymbol *BISubsecEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 882 MCSymbol *BIEnd = beginSymbolRecord(SymbolKind::S_BUILDINFO); 883 OS.AddComment("LF_BUILDINFO index"); 884 OS.emitInt32(BuildInfoIndex.getIndex()); 885 endSymbolRecord(BIEnd); 886 endCVSubsection(BISubsecEnd); 887 } 888 889 void CodeViewDebug::emitInlineeLinesSubsection() { 890 if (InlinedSubprograms.empty()) 891 return; 892 893 OS.AddComment("Inlinee lines subsection"); 894 MCSymbol *InlineEnd = beginCVSubsection(DebugSubsectionKind::InlineeLines); 895 896 // We emit the checksum info for files. This is used by debuggers to 897 // determine if a pdb matches the source before loading it. Visual Studio, 898 // for instance, will display a warning that the breakpoints are not valid if 899 // the pdb does not match the source. 900 OS.AddComment("Inlinee lines signature"); 901 OS.emitInt32(unsigned(InlineeLinesSignature::Normal)); 902 903 for (const DISubprogram *SP : InlinedSubprograms) { 904 assert(TypeIndices.count({SP, nullptr})); 905 TypeIndex InlineeIdx = TypeIndices[{SP, nullptr}]; 906 907 OS.AddBlankLine(); 908 unsigned FileId = maybeRecordFile(SP->getFile()); 909 OS.AddComment("Inlined function " + SP->getName() + " starts at " + 910 SP->getFilename() + Twine(':') + Twine(SP->getLine())); 911 OS.AddBlankLine(); 912 OS.AddComment("Type index of inlined function"); 913 OS.emitInt32(InlineeIdx.getIndex()); 914 OS.AddComment("Offset into filechecksum table"); 915 OS.emitCVFileChecksumOffsetDirective(FileId); 916 OS.AddComment("Starting line number"); 917 OS.emitInt32(SP->getLine()); 918 } 919 920 endCVSubsection(InlineEnd); 921 } 922 923 void CodeViewDebug::emitInlinedCallSite(const FunctionInfo &FI, 924 const DILocation *InlinedAt, 925 const InlineSite &Site) { 926 assert(TypeIndices.count({Site.Inlinee, nullptr})); 927 TypeIndex InlineeIdx = TypeIndices[{Site.Inlinee, nullptr}]; 928 929 // SymbolRecord 930 MCSymbol *InlineEnd = beginSymbolRecord(SymbolKind::S_INLINESITE); 931 932 OS.AddComment("PtrParent"); 933 OS.emitInt32(0); 934 OS.AddComment("PtrEnd"); 935 OS.emitInt32(0); 936 OS.AddComment("Inlinee type index"); 937 OS.emitInt32(InlineeIdx.getIndex()); 938 939 unsigned FileId = maybeRecordFile(Site.Inlinee->getFile()); 940 unsigned StartLineNum = Site.Inlinee->getLine(); 941 942 OS.emitCVInlineLinetableDirective(Site.SiteFuncId, FileId, StartLineNum, 943 FI.Begin, FI.End); 944 945 endSymbolRecord(InlineEnd); 946 947 emitLocalVariableList(FI, Site.InlinedLocals); 948 949 // Recurse on child inlined call sites before closing the scope. 950 for (const DILocation *ChildSite : Site.ChildSites) { 951 auto I = FI.InlineSites.find(ChildSite); 952 assert(I != FI.InlineSites.end() && 953 "child site not in function inline site map"); 954 emitInlinedCallSite(FI, ChildSite, I->second); 955 } 956 957 // Close the scope. 958 emitEndSymbolRecord(SymbolKind::S_INLINESITE_END); 959 } 960 961 void CodeViewDebug::switchToDebugSectionForSymbol(const MCSymbol *GVSym) { 962 // If we have a symbol, it may be in a section that is COMDAT. If so, find the 963 // comdat key. A section may be comdat because of -ffunction-sections or 964 // because it is comdat in the IR. 965 MCSectionCOFF *GVSec = 966 GVSym ? dyn_cast<MCSectionCOFF>(&GVSym->getSection()) : nullptr; 967 const MCSymbol *KeySym = GVSec ? GVSec->getCOMDATSymbol() : nullptr; 968 969 MCSectionCOFF *DebugSec = cast<MCSectionCOFF>( 970 Asm->getObjFileLowering().getCOFFDebugSymbolsSection()); 971 DebugSec = OS.getContext().getAssociativeCOFFSection(DebugSec, KeySym); 972 973 OS.SwitchSection(DebugSec); 974 975 // Emit the magic version number if this is the first time we've switched to 976 // this section. 977 if (ComdatDebugSections.insert(DebugSec).second) 978 emitCodeViewMagicVersion(); 979 } 980 981 // Emit an S_THUNK32/S_END symbol pair for a thunk routine. 982 // The only supported thunk ordinal is currently the standard type. 983 void CodeViewDebug::emitDebugInfoForThunk(const Function *GV, 984 FunctionInfo &FI, 985 const MCSymbol *Fn) { 986 std::string FuncName = 987 std::string(GlobalValue::dropLLVMManglingEscape(GV->getName())); 988 const ThunkOrdinal ordinal = ThunkOrdinal::Standard; // Only supported kind. 989 990 OS.AddComment("Symbol subsection for " + Twine(FuncName)); 991 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 992 993 // Emit S_THUNK32 994 MCSymbol *ThunkRecordEnd = beginSymbolRecord(SymbolKind::S_THUNK32); 995 OS.AddComment("PtrParent"); 996 OS.emitInt32(0); 997 OS.AddComment("PtrEnd"); 998 OS.emitInt32(0); 999 OS.AddComment("PtrNext"); 1000 OS.emitInt32(0); 1001 OS.AddComment("Thunk section relative address"); 1002 OS.EmitCOFFSecRel32(Fn, /*Offset=*/0); 1003 OS.AddComment("Thunk section index"); 1004 OS.EmitCOFFSectionIndex(Fn); 1005 OS.AddComment("Code size"); 1006 OS.emitAbsoluteSymbolDiff(FI.End, Fn, 2); 1007 OS.AddComment("Ordinal"); 1008 OS.emitInt8(unsigned(ordinal)); 1009 OS.AddComment("Function name"); 1010 emitNullTerminatedSymbolName(OS, FuncName); 1011 // Additional fields specific to the thunk ordinal would go here. 1012 endSymbolRecord(ThunkRecordEnd); 1013 1014 // Local variables/inlined routines are purposely omitted here. The point of 1015 // marking this as a thunk is so Visual Studio will NOT stop in this routine. 1016 1017 // Emit S_PROC_ID_END 1018 emitEndSymbolRecord(SymbolKind::S_PROC_ID_END); 1019 1020 endCVSubsection(SymbolsEnd); 1021 } 1022 1023 void CodeViewDebug::emitDebugInfoForFunction(const Function *GV, 1024 FunctionInfo &FI) { 1025 // For each function there is a separate subsection which holds the PC to 1026 // file:line table. 1027 const MCSymbol *Fn = Asm->getSymbol(GV); 1028 assert(Fn); 1029 1030 // Switch to the to a comdat section, if appropriate. 1031 switchToDebugSectionForSymbol(Fn); 1032 1033 std::string FuncName; 1034 auto *SP = GV->getSubprogram(); 1035 assert(SP); 1036 setCurrentSubprogram(SP); 1037 1038 if (SP->isThunk()) { 1039 emitDebugInfoForThunk(GV, FI, Fn); 1040 return; 1041 } 1042 1043 // If we have a display name, build the fully qualified name by walking the 1044 // chain of scopes. 1045 if (!SP->getName().empty()) 1046 FuncName = getFullyQualifiedName(SP->getScope(), SP->getName()); 1047 1048 // If our DISubprogram name is empty, use the mangled name. 1049 if (FuncName.empty()) 1050 FuncName = std::string(GlobalValue::dropLLVMManglingEscape(GV->getName())); 1051 1052 // Emit FPO data, but only on 32-bit x86. No other platforms use it. 1053 if (Triple(MMI->getModule()->getTargetTriple()).getArch() == Triple::x86) 1054 OS.EmitCVFPOData(Fn); 1055 1056 // Emit a symbol subsection, required by VS2012+ to find function boundaries. 1057 OS.AddComment("Symbol subsection for " + Twine(FuncName)); 1058 MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols); 1059 { 1060 SymbolKind ProcKind = GV->hasLocalLinkage() ? SymbolKind::S_LPROC32_ID 1061 : SymbolKind::S_GPROC32_ID; 1062 MCSymbol *ProcRecordEnd = beginSymbolRecord(ProcKind); 1063 1064 // These fields are filled in by tools like CVPACK which run after the fact. 1065 OS.AddComment("PtrParent"); 1066 OS.emitInt32(0); 1067 OS.AddComment("PtrEnd"); 1068 OS.emitInt32(0); 1069 OS.AddComment("PtrNext"); 1070 OS.emitInt32(0); 1071 // This is the important bit that tells the debugger where the function 1072 // code is located and what's its size: 1073 OS.AddComment("Code size"); 1074 OS.emitAbsoluteSymbolDiff(FI.End, Fn, 4); 1075 OS.AddComment("Offset after prologue"); 1076 OS.emitInt32(0); 1077 OS.AddComment("Offset before epilogue"); 1078 OS.emitInt32(0); 1079 OS.AddComment("Function type index"); 1080 OS.emitInt32(getFuncIdForSubprogram(GV->getSubprogram()).getIndex()); 1081 OS.AddComment("Function section relative address"); 1082 OS.EmitCOFFSecRel32(Fn, /*Offset=*/0); 1083 OS.AddComment("Function section index"); 1084 OS.EmitCOFFSectionIndex(Fn); 1085 OS.AddComment("Flags"); 1086 OS.emitInt8(0); 1087 // Emit the function display name as a null-terminated string. 1088 OS.AddComment("Function name"); 1089 // Truncate the name so we won't overflow the record length field. 1090 emitNullTerminatedSymbolName(OS, FuncName); 1091 endSymbolRecord(ProcRecordEnd); 1092 1093 MCSymbol *FrameProcEnd = beginSymbolRecord(SymbolKind::S_FRAMEPROC); 1094 // Subtract out the CSR size since MSVC excludes that and we include it. 1095 OS.AddComment("FrameSize"); 1096 OS.emitInt32(FI.FrameSize - FI.CSRSize); 1097 OS.AddComment("Padding"); 1098 OS.emitInt32(0); 1099 OS.AddComment("Offset of padding"); 1100 OS.emitInt32(0); 1101 OS.AddComment("Bytes of callee saved registers"); 1102 OS.emitInt32(FI.CSRSize); 1103 OS.AddComment("Exception handler offset"); 1104 OS.emitInt32(0); 1105 OS.AddComment("Exception handler section"); 1106 OS.emitInt16(0); 1107 OS.AddComment("Flags (defines frame register)"); 1108 OS.emitInt32(uint32_t(FI.FrameProcOpts)); 1109 endSymbolRecord(FrameProcEnd); 1110 1111 emitLocalVariableList(FI, FI.Locals); 1112 emitGlobalVariableList(FI.Globals); 1113 emitLexicalBlockList(FI.ChildBlocks, FI); 1114 1115 // Emit inlined call site information. Only emit functions inlined directly 1116 // into the parent function. We'll emit the other sites recursively as part 1117 // of their parent inline site. 1118 for (const DILocation *InlinedAt : FI.ChildSites) { 1119 auto I = FI.InlineSites.find(InlinedAt); 1120 assert(I != FI.InlineSites.end() && 1121 "child site not in function inline site map"); 1122 emitInlinedCallSite(FI, InlinedAt, I->second); 1123 } 1124 1125 for (auto Annot : FI.Annotations) { 1126 MCSymbol *Label = Annot.first; 1127 MDTuple *Strs = cast<MDTuple>(Annot.second); 1128 MCSymbol *AnnotEnd = beginSymbolRecord(SymbolKind::S_ANNOTATION); 1129 OS.EmitCOFFSecRel32(Label, /*Offset=*/0); 1130 // FIXME: Make sure we don't overflow the max record size. 1131 OS.EmitCOFFSectionIndex(Label); 1132 OS.emitInt16(Strs->getNumOperands()); 1133 for (Metadata *MD : Strs->operands()) { 1134 // MDStrings are null terminated, so we can do EmitBytes and get the 1135 // nice .asciz directive. 1136 StringRef Str = cast<MDString>(MD)->getString(); 1137 assert(Str.data()[Str.size()] == '\0' && "non-nullterminated MDString"); 1138 OS.emitBytes(StringRef(Str.data(), Str.size() + 1)); 1139 } 1140 endSymbolRecord(AnnotEnd); 1141 } 1142 1143 for (auto HeapAllocSite : FI.HeapAllocSites) { 1144 const MCSymbol *BeginLabel = std::get<0>(HeapAllocSite); 1145 const MCSymbol *EndLabel = std::get<1>(HeapAllocSite); 1146 const DIType *DITy = std::get<2>(HeapAllocSite); 1147 MCSymbol *HeapAllocEnd = beginSymbolRecord(SymbolKind::S_HEAPALLOCSITE); 1148 OS.AddComment("Call site offset"); 1149 OS.EmitCOFFSecRel32(BeginLabel, /*Offset=*/0); 1150 OS.AddComment("Call site section index"); 1151 OS.EmitCOFFSectionIndex(BeginLabel); 1152 OS.AddComment("Call instruction length"); 1153 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2); 1154 OS.AddComment("Type index"); 1155 OS.emitInt32(getCompleteTypeIndex(DITy).getIndex()); 1156 endSymbolRecord(HeapAllocEnd); 1157 } 1158 1159 if (SP != nullptr) 1160 emitDebugInfoForUDTs(LocalUDTs); 1161 1162 // We're done with this function. 1163 emitEndSymbolRecord(SymbolKind::S_PROC_ID_END); 1164 } 1165 endCVSubsection(SymbolsEnd); 1166 1167 // We have an assembler directive that takes care of the whole line table. 1168 OS.emitCVLinetableDirective(FI.FuncId, Fn, FI.End); 1169 } 1170 1171 CodeViewDebug::LocalVarDefRange 1172 CodeViewDebug::createDefRangeMem(uint16_t CVRegister, int Offset) { 1173 LocalVarDefRange DR; 1174 DR.InMemory = -1; 1175 DR.DataOffset = Offset; 1176 assert(DR.DataOffset == Offset && "truncation"); 1177 DR.IsSubfield = 0; 1178 DR.StructOffset = 0; 1179 DR.CVRegister = CVRegister; 1180 return DR; 1181 } 1182 1183 void CodeViewDebug::collectVariableInfoFromMFTable( 1184 DenseSet<InlinedEntity> &Processed) { 1185 const MachineFunction &MF = *Asm->MF; 1186 const TargetSubtargetInfo &TSI = MF.getSubtarget(); 1187 const TargetFrameLowering *TFI = TSI.getFrameLowering(); 1188 const TargetRegisterInfo *TRI = TSI.getRegisterInfo(); 1189 1190 for (const MachineFunction::VariableDbgInfo &VI : MF.getVariableDbgInfo()) { 1191 if (!VI.Var) 1192 continue; 1193 assert(VI.Var->isValidLocationForIntrinsic(VI.Loc) && 1194 "Expected inlined-at fields to agree"); 1195 1196 Processed.insert(InlinedEntity(VI.Var, VI.Loc->getInlinedAt())); 1197 LexicalScope *Scope = LScopes.findLexicalScope(VI.Loc); 1198 1199 // If variable scope is not found then skip this variable. 1200 if (!Scope) 1201 continue; 1202 1203 // If the variable has an attached offset expression, extract it. 1204 // FIXME: Try to handle DW_OP_deref as well. 1205 int64_t ExprOffset = 0; 1206 bool Deref = false; 1207 if (VI.Expr) { 1208 // If there is one DW_OP_deref element, use offset of 0 and keep going. 1209 if (VI.Expr->getNumElements() == 1 && 1210 VI.Expr->getElement(0) == llvm::dwarf::DW_OP_deref) 1211 Deref = true; 1212 else if (!VI.Expr->extractIfOffset(ExprOffset)) 1213 continue; 1214 } 1215 1216 // Get the frame register used and the offset. 1217 Register FrameReg; 1218 int FrameOffset = TFI->getFrameIndexReference(*Asm->MF, VI.Slot, FrameReg); 1219 uint16_t CVReg = TRI->getCodeViewRegNum(FrameReg); 1220 1221 // Calculate the label ranges. 1222 LocalVarDefRange DefRange = 1223 createDefRangeMem(CVReg, FrameOffset + ExprOffset); 1224 1225 for (const InsnRange &Range : Scope->getRanges()) { 1226 const MCSymbol *Begin = getLabelBeforeInsn(Range.first); 1227 const MCSymbol *End = getLabelAfterInsn(Range.second); 1228 End = End ? End : Asm->getFunctionEnd(); 1229 DefRange.Ranges.emplace_back(Begin, End); 1230 } 1231 1232 LocalVariable Var; 1233 Var.DIVar = VI.Var; 1234 Var.DefRanges.emplace_back(std::move(DefRange)); 1235 if (Deref) 1236 Var.UseReferenceType = true; 1237 1238 recordLocalVariable(std::move(Var), Scope); 1239 } 1240 } 1241 1242 static bool canUseReferenceType(const DbgVariableLocation &Loc) { 1243 return !Loc.LoadChain.empty() && Loc.LoadChain.back() == 0; 1244 } 1245 1246 static bool needsReferenceType(const DbgVariableLocation &Loc) { 1247 return Loc.LoadChain.size() == 2 && Loc.LoadChain.back() == 0; 1248 } 1249 1250 void CodeViewDebug::calculateRanges( 1251 LocalVariable &Var, const DbgValueHistoryMap::Entries &Entries) { 1252 const TargetRegisterInfo *TRI = Asm->MF->getSubtarget().getRegisterInfo(); 1253 1254 // Calculate the definition ranges. 1255 for (auto I = Entries.begin(), E = Entries.end(); I != E; ++I) { 1256 const auto &Entry = *I; 1257 if (!Entry.isDbgValue()) 1258 continue; 1259 const MachineInstr *DVInst = Entry.getInstr(); 1260 assert(DVInst->isDebugValue() && "Invalid History entry"); 1261 // FIXME: Find a way to represent constant variables, since they are 1262 // relatively common. 1263 Optional<DbgVariableLocation> Location = 1264 DbgVariableLocation::extractFromMachineInstruction(*DVInst); 1265 if (!Location) 1266 continue; 1267 1268 // CodeView can only express variables in register and variables in memory 1269 // at a constant offset from a register. However, for variables passed 1270 // indirectly by pointer, it is common for that pointer to be spilled to a 1271 // stack location. For the special case of one offseted load followed by a 1272 // zero offset load (a pointer spilled to the stack), we change the type of 1273 // the local variable from a value type to a reference type. This tricks the 1274 // debugger into doing the load for us. 1275 if (Var.UseReferenceType) { 1276 // We're using a reference type. Drop the last zero offset load. 1277 if (canUseReferenceType(*Location)) 1278 Location->LoadChain.pop_back(); 1279 else 1280 continue; 1281 } else if (needsReferenceType(*Location)) { 1282 // This location can't be expressed without switching to a reference type. 1283 // Start over using that. 1284 Var.UseReferenceType = true; 1285 Var.DefRanges.clear(); 1286 calculateRanges(Var, Entries); 1287 return; 1288 } 1289 1290 // We can only handle a register or an offseted load of a register. 1291 if (Location->Register == 0 || Location->LoadChain.size() > 1) 1292 continue; 1293 { 1294 LocalVarDefRange DR; 1295 DR.CVRegister = TRI->getCodeViewRegNum(Location->Register); 1296 DR.InMemory = !Location->LoadChain.empty(); 1297 DR.DataOffset = 1298 !Location->LoadChain.empty() ? Location->LoadChain.back() : 0; 1299 if (Location->FragmentInfo) { 1300 DR.IsSubfield = true; 1301 DR.StructOffset = Location->FragmentInfo->OffsetInBits / 8; 1302 } else { 1303 DR.IsSubfield = false; 1304 DR.StructOffset = 0; 1305 } 1306 1307 if (Var.DefRanges.empty() || 1308 Var.DefRanges.back().isDifferentLocation(DR)) { 1309 Var.DefRanges.emplace_back(std::move(DR)); 1310 } 1311 } 1312 1313 // Compute the label range. 1314 const MCSymbol *Begin = getLabelBeforeInsn(Entry.getInstr()); 1315 const MCSymbol *End; 1316 if (Entry.getEndIndex() != DbgValueHistoryMap::NoEntry) { 1317 auto &EndingEntry = Entries[Entry.getEndIndex()]; 1318 End = EndingEntry.isDbgValue() 1319 ? getLabelBeforeInsn(EndingEntry.getInstr()) 1320 : getLabelAfterInsn(EndingEntry.getInstr()); 1321 } else 1322 End = Asm->getFunctionEnd(); 1323 1324 // If the last range end is our begin, just extend the last range. 1325 // Otherwise make a new range. 1326 SmallVectorImpl<std::pair<const MCSymbol *, const MCSymbol *>> &R = 1327 Var.DefRanges.back().Ranges; 1328 if (!R.empty() && R.back().second == Begin) 1329 R.back().second = End; 1330 else 1331 R.emplace_back(Begin, End); 1332 1333 // FIXME: Do more range combining. 1334 } 1335 } 1336 1337 void CodeViewDebug::collectVariableInfo(const DISubprogram *SP) { 1338 DenseSet<InlinedEntity> Processed; 1339 // Grab the variable info that was squirreled away in the MMI side-table. 1340 collectVariableInfoFromMFTable(Processed); 1341 1342 for (const auto &I : DbgValues) { 1343 InlinedEntity IV = I.first; 1344 if (Processed.count(IV)) 1345 continue; 1346 const DILocalVariable *DIVar = cast<DILocalVariable>(IV.first); 1347 const DILocation *InlinedAt = IV.second; 1348 1349 // Instruction ranges, specifying where IV is accessible. 1350 const auto &Entries = I.second; 1351 1352 LexicalScope *Scope = nullptr; 1353 if (InlinedAt) 1354 Scope = LScopes.findInlinedScope(DIVar->getScope(), InlinedAt); 1355 else 1356 Scope = LScopes.findLexicalScope(DIVar->getScope()); 1357 // If variable scope is not found then skip this variable. 1358 if (!Scope) 1359 continue; 1360 1361 LocalVariable Var; 1362 Var.DIVar = DIVar; 1363 1364 calculateRanges(Var, Entries); 1365 recordLocalVariable(std::move(Var), Scope); 1366 } 1367 } 1368 1369 void CodeViewDebug::beginFunctionImpl(const MachineFunction *MF) { 1370 const TargetSubtargetInfo &TSI = MF->getSubtarget(); 1371 const TargetRegisterInfo *TRI = TSI.getRegisterInfo(); 1372 const MachineFrameInfo &MFI = MF->getFrameInfo(); 1373 const Function &GV = MF->getFunction(); 1374 auto Insertion = FnDebugInfo.insert({&GV, std::make_unique<FunctionInfo>()}); 1375 assert(Insertion.second && "function already has info"); 1376 CurFn = Insertion.first->second.get(); 1377 CurFn->FuncId = NextFuncId++; 1378 CurFn->Begin = Asm->getFunctionBegin(); 1379 1380 // The S_FRAMEPROC record reports the stack size, and how many bytes of 1381 // callee-saved registers were used. For targets that don't use a PUSH 1382 // instruction (AArch64), this will be zero. 1383 CurFn->CSRSize = MFI.getCVBytesOfCalleeSavedRegisters(); 1384 CurFn->FrameSize = MFI.getStackSize(); 1385 CurFn->OffsetAdjustment = MFI.getOffsetAdjustment(); 1386 CurFn->HasStackRealignment = TRI->needsStackRealignment(*MF); 1387 1388 // For this function S_FRAMEPROC record, figure out which codeview register 1389 // will be the frame pointer. 1390 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::None; // None. 1391 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::None; // None. 1392 if (CurFn->FrameSize > 0) { 1393 if (!TSI.getFrameLowering()->hasFP(*MF)) { 1394 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr; 1395 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::StackPtr; 1396 } else { 1397 // If there is an FP, parameters are always relative to it. 1398 CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::FramePtr; 1399 if (CurFn->HasStackRealignment) { 1400 // If the stack needs realignment, locals are relative to SP or VFRAME. 1401 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr; 1402 } else { 1403 // Otherwise, locals are relative to EBP, and we probably have VLAs or 1404 // other stack adjustments. 1405 CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::FramePtr; 1406 } 1407 } 1408 } 1409 1410 // Compute other frame procedure options. 1411 FrameProcedureOptions FPO = FrameProcedureOptions::None; 1412 if (MFI.hasVarSizedObjects()) 1413 FPO |= FrameProcedureOptions::HasAlloca; 1414 if (MF->exposesReturnsTwice()) 1415 FPO |= FrameProcedureOptions::HasSetJmp; 1416 // FIXME: Set HasLongJmp if we ever track that info. 1417 if (MF->hasInlineAsm()) 1418 FPO |= FrameProcedureOptions::HasInlineAssembly; 1419 if (GV.hasPersonalityFn()) { 1420 if (isAsynchronousEHPersonality( 1421 classifyEHPersonality(GV.getPersonalityFn()))) 1422 FPO |= FrameProcedureOptions::HasStructuredExceptionHandling; 1423 else 1424 FPO |= FrameProcedureOptions::HasExceptionHandling; 1425 } 1426 if (GV.hasFnAttribute(Attribute::InlineHint)) 1427 FPO |= FrameProcedureOptions::MarkedInline; 1428 if (GV.hasFnAttribute(Attribute::Naked)) 1429 FPO |= FrameProcedureOptions::Naked; 1430 if (MFI.hasStackProtectorIndex()) 1431 FPO |= FrameProcedureOptions::SecurityChecks; 1432 FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedLocalFramePtrReg) << 14U); 1433 FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedParamFramePtrReg) << 16U); 1434 if (Asm->TM.getOptLevel() != CodeGenOpt::None && 1435 !GV.hasOptSize() && !GV.hasOptNone()) 1436 FPO |= FrameProcedureOptions::OptimizedForSpeed; 1437 // FIXME: Set GuardCfg when it is implemented. 1438 CurFn->FrameProcOpts = FPO; 1439 1440 OS.EmitCVFuncIdDirective(CurFn->FuncId); 1441 1442 // Find the end of the function prolog. First known non-DBG_VALUE and 1443 // non-frame setup location marks the beginning of the function body. 1444 // FIXME: is there a simpler a way to do this? Can we just search 1445 // for the first instruction of the function, not the last of the prolog? 1446 DebugLoc PrologEndLoc; 1447 bool EmptyPrologue = true; 1448 for (const auto &MBB : *MF) { 1449 for (const auto &MI : MBB) { 1450 if (!MI.isMetaInstruction() && !MI.getFlag(MachineInstr::FrameSetup) && 1451 MI.getDebugLoc()) { 1452 PrologEndLoc = MI.getDebugLoc(); 1453 break; 1454 } else if (!MI.isMetaInstruction()) { 1455 EmptyPrologue = false; 1456 } 1457 } 1458 } 1459 1460 // Record beginning of function if we have a non-empty prologue. 1461 if (PrologEndLoc && !EmptyPrologue) { 1462 DebugLoc FnStartDL = PrologEndLoc.getFnDebugLoc(); 1463 maybeRecordLocation(FnStartDL, MF); 1464 } 1465 1466 // Find heap alloc sites and emit labels around them. 1467 for (const auto &MBB : *MF) { 1468 for (const auto &MI : MBB) { 1469 if (MI.getHeapAllocMarker()) { 1470 requestLabelBeforeInsn(&MI); 1471 requestLabelAfterInsn(&MI); 1472 } 1473 } 1474 } 1475 } 1476 1477 static bool shouldEmitUdt(const DIType *T) { 1478 if (!T) 1479 return false; 1480 1481 // MSVC does not emit UDTs for typedefs that are scoped to classes. 1482 if (T->getTag() == dwarf::DW_TAG_typedef) { 1483 if (DIScope *Scope = T->getScope()) { 1484 switch (Scope->getTag()) { 1485 case dwarf::DW_TAG_structure_type: 1486 case dwarf::DW_TAG_class_type: 1487 case dwarf::DW_TAG_union_type: 1488 return false; 1489 } 1490 } 1491 } 1492 1493 while (true) { 1494 if (!T || T->isForwardDecl()) 1495 return false; 1496 1497 const DIDerivedType *DT = dyn_cast<DIDerivedType>(T); 1498 if (!DT) 1499 return true; 1500 T = DT->getBaseType(); 1501 } 1502 return true; 1503 } 1504 1505 void CodeViewDebug::addToUDTs(const DIType *Ty) { 1506 // Don't record empty UDTs. 1507 if (Ty->getName().empty()) 1508 return; 1509 if (!shouldEmitUdt(Ty)) 1510 return; 1511 1512 SmallVector<StringRef, 5> ParentScopeNames; 1513 const DISubprogram *ClosestSubprogram = 1514 collectParentScopeNames(Ty->getScope(), ParentScopeNames); 1515 1516 std::string FullyQualifiedName = 1517 formatNestedName(ParentScopeNames, getPrettyScopeName(Ty)); 1518 1519 if (ClosestSubprogram == nullptr) { 1520 GlobalUDTs.emplace_back(std::move(FullyQualifiedName), Ty); 1521 } else if (ClosestSubprogram == CurrentSubprogram) { 1522 LocalUDTs.emplace_back(std::move(FullyQualifiedName), Ty); 1523 } 1524 1525 // TODO: What if the ClosestSubprogram is neither null or the current 1526 // subprogram? Currently, the UDT just gets dropped on the floor. 1527 // 1528 // The current behavior is not desirable. To get maximal fidelity, we would 1529 // need to perform all type translation before beginning emission of .debug$S 1530 // and then make LocalUDTs a member of FunctionInfo 1531 } 1532 1533 TypeIndex CodeViewDebug::lowerType(const DIType *Ty, const DIType *ClassTy) { 1534 // Generic dispatch for lowering an unknown type. 1535 switch (Ty->getTag()) { 1536 case dwarf::DW_TAG_array_type: 1537 return lowerTypeArray(cast<DICompositeType>(Ty)); 1538 case dwarf::DW_TAG_typedef: 1539 return lowerTypeAlias(cast<DIDerivedType>(Ty)); 1540 case dwarf::DW_TAG_base_type: 1541 return lowerTypeBasic(cast<DIBasicType>(Ty)); 1542 case dwarf::DW_TAG_pointer_type: 1543 if (cast<DIDerivedType>(Ty)->getName() == "__vtbl_ptr_type") 1544 return lowerTypeVFTableShape(cast<DIDerivedType>(Ty)); 1545 LLVM_FALLTHROUGH; 1546 case dwarf::DW_TAG_reference_type: 1547 case dwarf::DW_TAG_rvalue_reference_type: 1548 return lowerTypePointer(cast<DIDerivedType>(Ty)); 1549 case dwarf::DW_TAG_ptr_to_member_type: 1550 return lowerTypeMemberPointer(cast<DIDerivedType>(Ty)); 1551 case dwarf::DW_TAG_restrict_type: 1552 case dwarf::DW_TAG_const_type: 1553 case dwarf::DW_TAG_volatile_type: 1554 // TODO: add support for DW_TAG_atomic_type here 1555 return lowerTypeModifier(cast<DIDerivedType>(Ty)); 1556 case dwarf::DW_TAG_subroutine_type: 1557 if (ClassTy) { 1558 // The member function type of a member function pointer has no 1559 // ThisAdjustment. 1560 return lowerTypeMemberFunction(cast<DISubroutineType>(Ty), ClassTy, 1561 /*ThisAdjustment=*/0, 1562 /*IsStaticMethod=*/false); 1563 } 1564 return lowerTypeFunction(cast<DISubroutineType>(Ty)); 1565 case dwarf::DW_TAG_enumeration_type: 1566 return lowerTypeEnum(cast<DICompositeType>(Ty)); 1567 case dwarf::DW_TAG_class_type: 1568 case dwarf::DW_TAG_structure_type: 1569 return lowerTypeClass(cast<DICompositeType>(Ty)); 1570 case dwarf::DW_TAG_union_type: 1571 return lowerTypeUnion(cast<DICompositeType>(Ty)); 1572 case dwarf::DW_TAG_unspecified_type: 1573 if (Ty->getName() == "decltype(nullptr)") 1574 return TypeIndex::NullptrT(); 1575 return TypeIndex::None(); 1576 default: 1577 // Use the null type index. 1578 return TypeIndex(); 1579 } 1580 } 1581 1582 TypeIndex CodeViewDebug::lowerTypeAlias(const DIDerivedType *Ty) { 1583 TypeIndex UnderlyingTypeIndex = getTypeIndex(Ty->getBaseType()); 1584 StringRef TypeName = Ty->getName(); 1585 1586 addToUDTs(Ty); 1587 1588 if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::Int32Long) && 1589 TypeName == "HRESULT") 1590 return TypeIndex(SimpleTypeKind::HResult); 1591 if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::UInt16Short) && 1592 TypeName == "wchar_t") 1593 return TypeIndex(SimpleTypeKind::WideCharacter); 1594 1595 return UnderlyingTypeIndex; 1596 } 1597 1598 TypeIndex CodeViewDebug::lowerTypeArray(const DICompositeType *Ty) { 1599 const DIType *ElementType = Ty->getBaseType(); 1600 TypeIndex ElementTypeIndex = getTypeIndex(ElementType); 1601 // IndexType is size_t, which depends on the bitness of the target. 1602 TypeIndex IndexType = getPointerSizeInBytes() == 8 1603 ? TypeIndex(SimpleTypeKind::UInt64Quad) 1604 : TypeIndex(SimpleTypeKind::UInt32Long); 1605 1606 uint64_t ElementSize = getBaseTypeSize(ElementType) / 8; 1607 1608 // Add subranges to array type. 1609 DINodeArray Elements = Ty->getElements(); 1610 for (int i = Elements.size() - 1; i >= 0; --i) { 1611 const DINode *Element = Elements[i]; 1612 assert(Element->getTag() == dwarf::DW_TAG_subrange_type); 1613 1614 const DISubrange *Subrange = cast<DISubrange>(Element); 1615 assert(!Subrange->getRawLowerBound() && 1616 "codeview doesn't support subranges with lower bounds"); 1617 int64_t Count = -1; 1618 if (auto *CI = Subrange->getCount().dyn_cast<ConstantInt*>()) 1619 Count = CI->getSExtValue(); 1620 1621 // Forward declarations of arrays without a size and VLAs use a count of -1. 1622 // Emit a count of zero in these cases to match what MSVC does for arrays 1623 // without a size. MSVC doesn't support VLAs, so it's not clear what we 1624 // should do for them even if we could distinguish them. 1625 if (Count == -1) 1626 Count = 0; 1627 1628 // Update the element size and element type index for subsequent subranges. 1629 ElementSize *= Count; 1630 1631 // If this is the outermost array, use the size from the array. It will be 1632 // more accurate if we had a VLA or an incomplete element type size. 1633 uint64_t ArraySize = 1634 (i == 0 && ElementSize == 0) ? Ty->getSizeInBits() / 8 : ElementSize; 1635 1636 StringRef Name = (i == 0) ? Ty->getName() : ""; 1637 ArrayRecord AR(ElementTypeIndex, IndexType, ArraySize, Name); 1638 ElementTypeIndex = TypeTable.writeLeafType(AR); 1639 } 1640 1641 return ElementTypeIndex; 1642 } 1643 1644 TypeIndex CodeViewDebug::lowerTypeBasic(const DIBasicType *Ty) { 1645 TypeIndex Index; 1646 dwarf::TypeKind Kind; 1647 uint32_t ByteSize; 1648 1649 Kind = static_cast<dwarf::TypeKind>(Ty->getEncoding()); 1650 ByteSize = Ty->getSizeInBits() / 8; 1651 1652 SimpleTypeKind STK = SimpleTypeKind::None; 1653 switch (Kind) { 1654 case dwarf::DW_ATE_address: 1655 // FIXME: Translate 1656 break; 1657 case dwarf::DW_ATE_boolean: 1658 switch (ByteSize) { 1659 case 1: STK = SimpleTypeKind::Boolean8; break; 1660 case 2: STK = SimpleTypeKind::Boolean16; break; 1661 case 4: STK = SimpleTypeKind::Boolean32; break; 1662 case 8: STK = SimpleTypeKind::Boolean64; break; 1663 case 16: STK = SimpleTypeKind::Boolean128; break; 1664 } 1665 break; 1666 case dwarf::DW_ATE_complex_float: 1667 switch (ByteSize) { 1668 case 2: STK = SimpleTypeKind::Complex16; break; 1669 case 4: STK = SimpleTypeKind::Complex32; break; 1670 case 8: STK = SimpleTypeKind::Complex64; break; 1671 case 10: STK = SimpleTypeKind::Complex80; break; 1672 case 16: STK = SimpleTypeKind::Complex128; break; 1673 } 1674 break; 1675 case dwarf::DW_ATE_float: 1676 switch (ByteSize) { 1677 case 2: STK = SimpleTypeKind::Float16; break; 1678 case 4: STK = SimpleTypeKind::Float32; break; 1679 case 6: STK = SimpleTypeKind::Float48; break; 1680 case 8: STK = SimpleTypeKind::Float64; break; 1681 case 10: STK = SimpleTypeKind::Float80; break; 1682 case 16: STK = SimpleTypeKind::Float128; break; 1683 } 1684 break; 1685 case dwarf::DW_ATE_signed: 1686 switch (ByteSize) { 1687 case 1: STK = SimpleTypeKind::SignedCharacter; break; 1688 case 2: STK = SimpleTypeKind::Int16Short; break; 1689 case 4: STK = SimpleTypeKind::Int32; break; 1690 case 8: STK = SimpleTypeKind::Int64Quad; break; 1691 case 16: STK = SimpleTypeKind::Int128Oct; break; 1692 } 1693 break; 1694 case dwarf::DW_ATE_unsigned: 1695 switch (ByteSize) { 1696 case 1: STK = SimpleTypeKind::UnsignedCharacter; break; 1697 case 2: STK = SimpleTypeKind::UInt16Short; break; 1698 case 4: STK = SimpleTypeKind::UInt32; break; 1699 case 8: STK = SimpleTypeKind::UInt64Quad; break; 1700 case 16: STK = SimpleTypeKind::UInt128Oct; break; 1701 } 1702 break; 1703 case dwarf::DW_ATE_UTF: 1704 switch (ByteSize) { 1705 case 2: STK = SimpleTypeKind::Character16; break; 1706 case 4: STK = SimpleTypeKind::Character32; break; 1707 } 1708 break; 1709 case dwarf::DW_ATE_signed_char: 1710 if (ByteSize == 1) 1711 STK = SimpleTypeKind::SignedCharacter; 1712 break; 1713 case dwarf::DW_ATE_unsigned_char: 1714 if (ByteSize == 1) 1715 STK = SimpleTypeKind::UnsignedCharacter; 1716 break; 1717 default: 1718 break; 1719 } 1720 1721 // Apply some fixups based on the source-level type name. 1722 if (STK == SimpleTypeKind::Int32 && Ty->getName() == "long int") 1723 STK = SimpleTypeKind::Int32Long; 1724 if (STK == SimpleTypeKind::UInt32 && Ty->getName() == "long unsigned int") 1725 STK = SimpleTypeKind::UInt32Long; 1726 if (STK == SimpleTypeKind::UInt16Short && 1727 (Ty->getName() == "wchar_t" || Ty->getName() == "__wchar_t")) 1728 STK = SimpleTypeKind::WideCharacter; 1729 if ((STK == SimpleTypeKind::SignedCharacter || 1730 STK == SimpleTypeKind::UnsignedCharacter) && 1731 Ty->getName() == "char") 1732 STK = SimpleTypeKind::NarrowCharacter; 1733 1734 return TypeIndex(STK); 1735 } 1736 1737 TypeIndex CodeViewDebug::lowerTypePointer(const DIDerivedType *Ty, 1738 PointerOptions PO) { 1739 TypeIndex PointeeTI = getTypeIndex(Ty->getBaseType()); 1740 1741 // Pointers to simple types without any options can use SimpleTypeMode, rather 1742 // than having a dedicated pointer type record. 1743 if (PointeeTI.isSimple() && PO == PointerOptions::None && 1744 PointeeTI.getSimpleMode() == SimpleTypeMode::Direct && 1745 Ty->getTag() == dwarf::DW_TAG_pointer_type) { 1746 SimpleTypeMode Mode = Ty->getSizeInBits() == 64 1747 ? SimpleTypeMode::NearPointer64 1748 : SimpleTypeMode::NearPointer32; 1749 return TypeIndex(PointeeTI.getSimpleKind(), Mode); 1750 } 1751 1752 PointerKind PK = 1753 Ty->getSizeInBits() == 64 ? PointerKind::Near64 : PointerKind::Near32; 1754 PointerMode PM = PointerMode::Pointer; 1755 switch (Ty->getTag()) { 1756 default: llvm_unreachable("not a pointer tag type"); 1757 case dwarf::DW_TAG_pointer_type: 1758 PM = PointerMode::Pointer; 1759 break; 1760 case dwarf::DW_TAG_reference_type: 1761 PM = PointerMode::LValueReference; 1762 break; 1763 case dwarf::DW_TAG_rvalue_reference_type: 1764 PM = PointerMode::RValueReference; 1765 break; 1766 } 1767 1768 if (Ty->isObjectPointer()) 1769 PO |= PointerOptions::Const; 1770 1771 PointerRecord PR(PointeeTI, PK, PM, PO, Ty->getSizeInBits() / 8); 1772 return TypeTable.writeLeafType(PR); 1773 } 1774 1775 static PointerToMemberRepresentation 1776 translatePtrToMemberRep(unsigned SizeInBytes, bool IsPMF, unsigned Flags) { 1777 // SizeInBytes being zero generally implies that the member pointer type was 1778 // incomplete, which can happen if it is part of a function prototype. In this 1779 // case, use the unknown model instead of the general model. 1780 if (IsPMF) { 1781 switch (Flags & DINode::FlagPtrToMemberRep) { 1782 case 0: 1783 return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown 1784 : PointerToMemberRepresentation::GeneralFunction; 1785 case DINode::FlagSingleInheritance: 1786 return PointerToMemberRepresentation::SingleInheritanceFunction; 1787 case DINode::FlagMultipleInheritance: 1788 return PointerToMemberRepresentation::MultipleInheritanceFunction; 1789 case DINode::FlagVirtualInheritance: 1790 return PointerToMemberRepresentation::VirtualInheritanceFunction; 1791 } 1792 } else { 1793 switch (Flags & DINode::FlagPtrToMemberRep) { 1794 case 0: 1795 return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown 1796 : PointerToMemberRepresentation::GeneralData; 1797 case DINode::FlagSingleInheritance: 1798 return PointerToMemberRepresentation::SingleInheritanceData; 1799 case DINode::FlagMultipleInheritance: 1800 return PointerToMemberRepresentation::MultipleInheritanceData; 1801 case DINode::FlagVirtualInheritance: 1802 return PointerToMemberRepresentation::VirtualInheritanceData; 1803 } 1804 } 1805 llvm_unreachable("invalid ptr to member representation"); 1806 } 1807 1808 TypeIndex CodeViewDebug::lowerTypeMemberPointer(const DIDerivedType *Ty, 1809 PointerOptions PO) { 1810 assert(Ty->getTag() == dwarf::DW_TAG_ptr_to_member_type); 1811 bool IsPMF = isa<DISubroutineType>(Ty->getBaseType()); 1812 TypeIndex ClassTI = getTypeIndex(Ty->getClassType()); 1813 TypeIndex PointeeTI = 1814 getTypeIndex(Ty->getBaseType(), IsPMF ? Ty->getClassType() : nullptr); 1815 PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64 1816 : PointerKind::Near32; 1817 PointerMode PM = IsPMF ? PointerMode::PointerToMemberFunction 1818 : PointerMode::PointerToDataMember; 1819 1820 assert(Ty->getSizeInBits() / 8 <= 0xff && "pointer size too big"); 1821 uint8_t SizeInBytes = Ty->getSizeInBits() / 8; 1822 MemberPointerInfo MPI( 1823 ClassTI, translatePtrToMemberRep(SizeInBytes, IsPMF, Ty->getFlags())); 1824 PointerRecord PR(PointeeTI, PK, PM, PO, SizeInBytes, MPI); 1825 return TypeTable.writeLeafType(PR); 1826 } 1827 1828 /// Given a DWARF calling convention, get the CodeView equivalent. If we don't 1829 /// have a translation, use the NearC convention. 1830 static CallingConvention dwarfCCToCodeView(unsigned DwarfCC) { 1831 switch (DwarfCC) { 1832 case dwarf::DW_CC_normal: return CallingConvention::NearC; 1833 case dwarf::DW_CC_BORLAND_msfastcall: return CallingConvention::NearFast; 1834 case dwarf::DW_CC_BORLAND_thiscall: return CallingConvention::ThisCall; 1835 case dwarf::DW_CC_BORLAND_stdcall: return CallingConvention::NearStdCall; 1836 case dwarf::DW_CC_BORLAND_pascal: return CallingConvention::NearPascal; 1837 case dwarf::DW_CC_LLVM_vectorcall: return CallingConvention::NearVector; 1838 } 1839 return CallingConvention::NearC; 1840 } 1841 1842 TypeIndex CodeViewDebug::lowerTypeModifier(const DIDerivedType *Ty) { 1843 ModifierOptions Mods = ModifierOptions::None; 1844 PointerOptions PO = PointerOptions::None; 1845 bool IsModifier = true; 1846 const DIType *BaseTy = Ty; 1847 while (IsModifier && BaseTy) { 1848 // FIXME: Need to add DWARF tags for __unaligned and _Atomic 1849 switch (BaseTy->getTag()) { 1850 case dwarf::DW_TAG_const_type: 1851 Mods |= ModifierOptions::Const; 1852 PO |= PointerOptions::Const; 1853 break; 1854 case dwarf::DW_TAG_volatile_type: 1855 Mods |= ModifierOptions::Volatile; 1856 PO |= PointerOptions::Volatile; 1857 break; 1858 case dwarf::DW_TAG_restrict_type: 1859 // Only pointer types be marked with __restrict. There is no known flag 1860 // for __restrict in LF_MODIFIER records. 1861 PO |= PointerOptions::Restrict; 1862 break; 1863 default: 1864 IsModifier = false; 1865 break; 1866 } 1867 if (IsModifier) 1868 BaseTy = cast<DIDerivedType>(BaseTy)->getBaseType(); 1869 } 1870 1871 // Check if the inner type will use an LF_POINTER record. If so, the 1872 // qualifiers will go in the LF_POINTER record. This comes up for types like 1873 // 'int *const' and 'int *__restrict', not the more common cases like 'const 1874 // char *'. 1875 if (BaseTy) { 1876 switch (BaseTy->getTag()) { 1877 case dwarf::DW_TAG_pointer_type: 1878 case dwarf::DW_TAG_reference_type: 1879 case dwarf::DW_TAG_rvalue_reference_type: 1880 return lowerTypePointer(cast<DIDerivedType>(BaseTy), PO); 1881 case dwarf::DW_TAG_ptr_to_member_type: 1882 return lowerTypeMemberPointer(cast<DIDerivedType>(BaseTy), PO); 1883 default: 1884 break; 1885 } 1886 } 1887 1888 TypeIndex ModifiedTI = getTypeIndex(BaseTy); 1889 1890 // Return the base type index if there aren't any modifiers. For example, the 1891 // metadata could contain restrict wrappers around non-pointer types. 1892 if (Mods == ModifierOptions::None) 1893 return ModifiedTI; 1894 1895 ModifierRecord MR(ModifiedTI, Mods); 1896 return TypeTable.writeLeafType(MR); 1897 } 1898 1899 TypeIndex CodeViewDebug::lowerTypeFunction(const DISubroutineType *Ty) { 1900 SmallVector<TypeIndex, 8> ReturnAndArgTypeIndices; 1901 for (const DIType *ArgType : Ty->getTypeArray()) 1902 ReturnAndArgTypeIndices.push_back(getTypeIndex(ArgType)); 1903 1904 // MSVC uses type none for variadic argument. 1905 if (ReturnAndArgTypeIndices.size() > 1 && 1906 ReturnAndArgTypeIndices.back() == TypeIndex::Void()) { 1907 ReturnAndArgTypeIndices.back() = TypeIndex::None(); 1908 } 1909 TypeIndex ReturnTypeIndex = TypeIndex::Void(); 1910 ArrayRef<TypeIndex> ArgTypeIndices = None; 1911 if (!ReturnAndArgTypeIndices.empty()) { 1912 auto ReturnAndArgTypesRef = makeArrayRef(ReturnAndArgTypeIndices); 1913 ReturnTypeIndex = ReturnAndArgTypesRef.front(); 1914 ArgTypeIndices = ReturnAndArgTypesRef.drop_front(); 1915 } 1916 1917 ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices); 1918 TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec); 1919 1920 CallingConvention CC = dwarfCCToCodeView(Ty->getCC()); 1921 1922 FunctionOptions FO = getFunctionOptions(Ty); 1923 ProcedureRecord Procedure(ReturnTypeIndex, CC, FO, ArgTypeIndices.size(), 1924 ArgListIndex); 1925 return TypeTable.writeLeafType(Procedure); 1926 } 1927 1928 TypeIndex CodeViewDebug::lowerTypeMemberFunction(const DISubroutineType *Ty, 1929 const DIType *ClassTy, 1930 int ThisAdjustment, 1931 bool IsStaticMethod, 1932 FunctionOptions FO) { 1933 // Lower the containing class type. 1934 TypeIndex ClassType = getTypeIndex(ClassTy); 1935 1936 DITypeRefArray ReturnAndArgs = Ty->getTypeArray(); 1937 1938 unsigned Index = 0; 1939 SmallVector<TypeIndex, 8> ArgTypeIndices; 1940 TypeIndex ReturnTypeIndex = TypeIndex::Void(); 1941 if (ReturnAndArgs.size() > Index) { 1942 ReturnTypeIndex = getTypeIndex(ReturnAndArgs[Index++]); 1943 } 1944 1945 // If the first argument is a pointer type and this isn't a static method, 1946 // treat it as the special 'this' parameter, which is encoded separately from 1947 // the arguments. 1948 TypeIndex ThisTypeIndex; 1949 if (!IsStaticMethod && ReturnAndArgs.size() > Index) { 1950 if (const DIDerivedType *PtrTy = 1951 dyn_cast_or_null<DIDerivedType>(ReturnAndArgs[Index])) { 1952 if (PtrTy->getTag() == dwarf::DW_TAG_pointer_type) { 1953 ThisTypeIndex = getTypeIndexForThisPtr(PtrTy, Ty); 1954 Index++; 1955 } 1956 } 1957 } 1958 1959 while (Index < ReturnAndArgs.size()) 1960 ArgTypeIndices.push_back(getTypeIndex(ReturnAndArgs[Index++])); 1961 1962 // MSVC uses type none for variadic argument. 1963 if (!ArgTypeIndices.empty() && ArgTypeIndices.back() == TypeIndex::Void()) 1964 ArgTypeIndices.back() = TypeIndex::None(); 1965 1966 ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices); 1967 TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec); 1968 1969 CallingConvention CC = dwarfCCToCodeView(Ty->getCC()); 1970 1971 MemberFunctionRecord MFR(ReturnTypeIndex, ClassType, ThisTypeIndex, CC, FO, 1972 ArgTypeIndices.size(), ArgListIndex, ThisAdjustment); 1973 return TypeTable.writeLeafType(MFR); 1974 } 1975 1976 TypeIndex CodeViewDebug::lowerTypeVFTableShape(const DIDerivedType *Ty) { 1977 unsigned VSlotCount = 1978 Ty->getSizeInBits() / (8 * Asm->MAI->getCodePointerSize()); 1979 SmallVector<VFTableSlotKind, 4> Slots(VSlotCount, VFTableSlotKind::Near); 1980 1981 VFTableShapeRecord VFTSR(Slots); 1982 return TypeTable.writeLeafType(VFTSR); 1983 } 1984 1985 static MemberAccess translateAccessFlags(unsigned RecordTag, unsigned Flags) { 1986 switch (Flags & DINode::FlagAccessibility) { 1987 case DINode::FlagPrivate: return MemberAccess::Private; 1988 case DINode::FlagPublic: return MemberAccess::Public; 1989 case DINode::FlagProtected: return MemberAccess::Protected; 1990 case 0: 1991 // If there was no explicit access control, provide the default for the tag. 1992 return RecordTag == dwarf::DW_TAG_class_type ? MemberAccess::Private 1993 : MemberAccess::Public; 1994 } 1995 llvm_unreachable("access flags are exclusive"); 1996 } 1997 1998 static MethodOptions translateMethodOptionFlags(const DISubprogram *SP) { 1999 if (SP->isArtificial()) 2000 return MethodOptions::CompilerGenerated; 2001 2002 // FIXME: Handle other MethodOptions. 2003 2004 return MethodOptions::None; 2005 } 2006 2007 static MethodKind translateMethodKindFlags(const DISubprogram *SP, 2008 bool Introduced) { 2009 if (SP->getFlags() & DINode::FlagStaticMember) 2010 return MethodKind::Static; 2011 2012 switch (SP->getVirtuality()) { 2013 case dwarf::DW_VIRTUALITY_none: 2014 break; 2015 case dwarf::DW_VIRTUALITY_virtual: 2016 return Introduced ? MethodKind::IntroducingVirtual : MethodKind::Virtual; 2017 case dwarf::DW_VIRTUALITY_pure_virtual: 2018 return Introduced ? MethodKind::PureIntroducingVirtual 2019 : MethodKind::PureVirtual; 2020 default: 2021 llvm_unreachable("unhandled virtuality case"); 2022 } 2023 2024 return MethodKind::Vanilla; 2025 } 2026 2027 static TypeRecordKind getRecordKind(const DICompositeType *Ty) { 2028 switch (Ty->getTag()) { 2029 case dwarf::DW_TAG_class_type: return TypeRecordKind::Class; 2030 case dwarf::DW_TAG_structure_type: return TypeRecordKind::Struct; 2031 } 2032 llvm_unreachable("unexpected tag"); 2033 } 2034 2035 /// Return ClassOptions that should be present on both the forward declaration 2036 /// and the defintion of a tag type. 2037 static ClassOptions getCommonClassOptions(const DICompositeType *Ty) { 2038 ClassOptions CO = ClassOptions::None; 2039 2040 // MSVC always sets this flag, even for local types. Clang doesn't always 2041 // appear to give every type a linkage name, which may be problematic for us. 2042 // FIXME: Investigate the consequences of not following them here. 2043 if (!Ty->getIdentifier().empty()) 2044 CO |= ClassOptions::HasUniqueName; 2045 2046 // Put the Nested flag on a type if it appears immediately inside a tag type. 2047 // Do not walk the scope chain. Do not attempt to compute ContainsNestedClass 2048 // here. That flag is only set on definitions, and not forward declarations. 2049 const DIScope *ImmediateScope = Ty->getScope(); 2050 if (ImmediateScope && isa<DICompositeType>(ImmediateScope)) 2051 CO |= ClassOptions::Nested; 2052 2053 // Put the Scoped flag on function-local types. MSVC puts this flag for enum 2054 // type only when it has an immediate function scope. Clang never puts enums 2055 // inside DILexicalBlock scopes. Enum types, as generated by clang, are 2056 // always in function, class, or file scopes. 2057 if (Ty->getTag() == dwarf::DW_TAG_enumeration_type) { 2058 if (ImmediateScope && isa<DISubprogram>(ImmediateScope)) 2059 CO |= ClassOptions::Scoped; 2060 } else { 2061 for (const DIScope *Scope = ImmediateScope; Scope != nullptr; 2062 Scope = Scope->getScope()) { 2063 if (isa<DISubprogram>(Scope)) { 2064 CO |= ClassOptions::Scoped; 2065 break; 2066 } 2067 } 2068 } 2069 2070 return CO; 2071 } 2072 2073 void CodeViewDebug::addUDTSrcLine(const DIType *Ty, TypeIndex TI) { 2074 switch (Ty->getTag()) { 2075 case dwarf::DW_TAG_class_type: 2076 case dwarf::DW_TAG_structure_type: 2077 case dwarf::DW_TAG_union_type: 2078 case dwarf::DW_TAG_enumeration_type: 2079 break; 2080 default: 2081 return; 2082 } 2083 2084 if (const auto *File = Ty->getFile()) { 2085 StringIdRecord SIDR(TypeIndex(0x0), getFullFilepath(File)); 2086 TypeIndex SIDI = TypeTable.writeLeafType(SIDR); 2087 2088 UdtSourceLineRecord USLR(TI, SIDI, Ty->getLine()); 2089 TypeTable.writeLeafType(USLR); 2090 } 2091 } 2092 2093 TypeIndex CodeViewDebug::lowerTypeEnum(const DICompositeType *Ty) { 2094 ClassOptions CO = getCommonClassOptions(Ty); 2095 TypeIndex FTI; 2096 unsigned EnumeratorCount = 0; 2097 2098 if (Ty->isForwardDecl()) { 2099 CO |= ClassOptions::ForwardReference; 2100 } else { 2101 ContinuationRecordBuilder ContinuationBuilder; 2102 ContinuationBuilder.begin(ContinuationRecordKind::FieldList); 2103 for (const DINode *Element : Ty->getElements()) { 2104 // We assume that the frontend provides all members in source declaration 2105 // order, which is what MSVC does. 2106 if (auto *Enumerator = dyn_cast_or_null<DIEnumerator>(Element)) { 2107 EnumeratorRecord ER(MemberAccess::Public, 2108 APSInt(Enumerator->getValue(), true), 2109 Enumerator->getName()); 2110 ContinuationBuilder.writeMemberType(ER); 2111 EnumeratorCount++; 2112 } 2113 } 2114 FTI = TypeTable.insertRecord(ContinuationBuilder); 2115 } 2116 2117 std::string FullName = getFullyQualifiedName(Ty); 2118 2119 EnumRecord ER(EnumeratorCount, CO, FTI, FullName, Ty->getIdentifier(), 2120 getTypeIndex(Ty->getBaseType())); 2121 TypeIndex EnumTI = TypeTable.writeLeafType(ER); 2122 2123 addUDTSrcLine(Ty, EnumTI); 2124 2125 return EnumTI; 2126 } 2127 2128 //===----------------------------------------------------------------------===// 2129 // ClassInfo 2130 //===----------------------------------------------------------------------===// 2131 2132 struct llvm::ClassInfo { 2133 struct MemberInfo { 2134 const DIDerivedType *MemberTypeNode; 2135 uint64_t BaseOffset; 2136 }; 2137 // [MemberInfo] 2138 using MemberList = std::vector<MemberInfo>; 2139 2140 using MethodsList = TinyPtrVector<const DISubprogram *>; 2141 // MethodName -> MethodsList 2142 using MethodsMap = MapVector<MDString *, MethodsList>; 2143 2144 /// Base classes. 2145 std::vector<const DIDerivedType *> Inheritance; 2146 2147 /// Direct members. 2148 MemberList Members; 2149 // Direct overloaded methods gathered by name. 2150 MethodsMap Methods; 2151 2152 TypeIndex VShapeTI; 2153 2154 std::vector<const DIType *> NestedTypes; 2155 }; 2156 2157 void CodeViewDebug::clear() { 2158 assert(CurFn == nullptr); 2159 FileIdMap.clear(); 2160 FnDebugInfo.clear(); 2161 FileToFilepathMap.clear(); 2162 LocalUDTs.clear(); 2163 GlobalUDTs.clear(); 2164 TypeIndices.clear(); 2165 CompleteTypeIndices.clear(); 2166 ScopeGlobals.clear(); 2167 } 2168 2169 void CodeViewDebug::collectMemberInfo(ClassInfo &Info, 2170 const DIDerivedType *DDTy) { 2171 if (!DDTy->getName().empty()) { 2172 Info.Members.push_back({DDTy, 0}); 2173 return; 2174 } 2175 2176 // An unnamed member may represent a nested struct or union. Attempt to 2177 // interpret the unnamed member as a DICompositeType possibly wrapped in 2178 // qualifier types. Add all the indirect fields to the current record if that 2179 // succeeds, and drop the member if that fails. 2180 assert((DDTy->getOffsetInBits() % 8) == 0 && "Unnamed bitfield member!"); 2181 uint64_t Offset = DDTy->getOffsetInBits(); 2182 const DIType *Ty = DDTy->getBaseType(); 2183 bool FullyResolved = false; 2184 while (!FullyResolved) { 2185 switch (Ty->getTag()) { 2186 case dwarf::DW_TAG_const_type: 2187 case dwarf::DW_TAG_volatile_type: 2188 // FIXME: we should apply the qualifier types to the indirect fields 2189 // rather than dropping them. 2190 Ty = cast<DIDerivedType>(Ty)->getBaseType(); 2191 break; 2192 default: 2193 FullyResolved = true; 2194 break; 2195 } 2196 } 2197 2198 const DICompositeType *DCTy = dyn_cast<DICompositeType>(Ty); 2199 if (!DCTy) 2200 return; 2201 2202 ClassInfo NestedInfo = collectClassInfo(DCTy); 2203 for (const ClassInfo::MemberInfo &IndirectField : NestedInfo.Members) 2204 Info.Members.push_back( 2205 {IndirectField.MemberTypeNode, IndirectField.BaseOffset + Offset}); 2206 } 2207 2208 ClassInfo CodeViewDebug::collectClassInfo(const DICompositeType *Ty) { 2209 ClassInfo Info; 2210 // Add elements to structure type. 2211 DINodeArray Elements = Ty->getElements(); 2212 for (auto *Element : Elements) { 2213 // We assume that the frontend provides all members in source declaration 2214 // order, which is what MSVC does. 2215 if (!Element) 2216 continue; 2217 if (auto *SP = dyn_cast<DISubprogram>(Element)) { 2218 Info.Methods[SP->getRawName()].push_back(SP); 2219 } else if (auto *DDTy = dyn_cast<DIDerivedType>(Element)) { 2220 if (DDTy->getTag() == dwarf::DW_TAG_member) { 2221 collectMemberInfo(Info, DDTy); 2222 } else if (DDTy->getTag() == dwarf::DW_TAG_inheritance) { 2223 Info.Inheritance.push_back(DDTy); 2224 } else if (DDTy->getTag() == dwarf::DW_TAG_pointer_type && 2225 DDTy->getName() == "__vtbl_ptr_type") { 2226 Info.VShapeTI = getTypeIndex(DDTy); 2227 } else if (DDTy->getTag() == dwarf::DW_TAG_typedef) { 2228 Info.NestedTypes.push_back(DDTy); 2229 } else if (DDTy->getTag() == dwarf::DW_TAG_friend) { 2230 // Ignore friend members. It appears that MSVC emitted info about 2231 // friends in the past, but modern versions do not. 2232 } 2233 } else if (auto *Composite = dyn_cast<DICompositeType>(Element)) { 2234 Info.NestedTypes.push_back(Composite); 2235 } 2236 // Skip other unrecognized kinds of elements. 2237 } 2238 return Info; 2239 } 2240 2241 static bool shouldAlwaysEmitCompleteClassType(const DICompositeType *Ty) { 2242 // This routine is used by lowerTypeClass and lowerTypeUnion to determine 2243 // if a complete type should be emitted instead of a forward reference. 2244 return Ty->getName().empty() && Ty->getIdentifier().empty() && 2245 !Ty->isForwardDecl(); 2246 } 2247 2248 TypeIndex CodeViewDebug::lowerTypeClass(const DICompositeType *Ty) { 2249 // Emit the complete type for unnamed structs. C++ classes with methods 2250 // which have a circular reference back to the class type are expected to 2251 // be named by the front-end and should not be "unnamed". C unnamed 2252 // structs should not have circular references. 2253 if (shouldAlwaysEmitCompleteClassType(Ty)) { 2254 // If this unnamed complete type is already in the process of being defined 2255 // then the description of the type is malformed and cannot be emitted 2256 // into CodeView correctly so report a fatal error. 2257 auto I = CompleteTypeIndices.find(Ty); 2258 if (I != CompleteTypeIndices.end() && I->second == TypeIndex()) 2259 report_fatal_error("cannot debug circular reference to unnamed type"); 2260 return getCompleteTypeIndex(Ty); 2261 } 2262 2263 // First, construct the forward decl. Don't look into Ty to compute the 2264 // forward decl options, since it might not be available in all TUs. 2265 TypeRecordKind Kind = getRecordKind(Ty); 2266 ClassOptions CO = 2267 ClassOptions::ForwardReference | getCommonClassOptions(Ty); 2268 std::string FullName = getFullyQualifiedName(Ty); 2269 ClassRecord CR(Kind, 0, CO, TypeIndex(), TypeIndex(), TypeIndex(), 0, 2270 FullName, Ty->getIdentifier()); 2271 TypeIndex FwdDeclTI = TypeTable.writeLeafType(CR); 2272 if (!Ty->isForwardDecl()) 2273 DeferredCompleteTypes.push_back(Ty); 2274 return FwdDeclTI; 2275 } 2276 2277 TypeIndex CodeViewDebug::lowerCompleteTypeClass(const DICompositeType *Ty) { 2278 // Construct the field list and complete type record. 2279 TypeRecordKind Kind = getRecordKind(Ty); 2280 ClassOptions CO = getCommonClassOptions(Ty); 2281 TypeIndex FieldTI; 2282 TypeIndex VShapeTI; 2283 unsigned FieldCount; 2284 bool ContainsNestedClass; 2285 std::tie(FieldTI, VShapeTI, FieldCount, ContainsNestedClass) = 2286 lowerRecordFieldList(Ty); 2287 2288 if (ContainsNestedClass) 2289 CO |= ClassOptions::ContainsNestedClass; 2290 2291 // MSVC appears to set this flag by searching any destructor or method with 2292 // FunctionOptions::Constructor among the emitted members. Clang AST has all 2293 // the members, however special member functions are not yet emitted into 2294 // debug information. For now checking a class's non-triviality seems enough. 2295 // FIXME: not true for a nested unnamed struct. 2296 if (isNonTrivial(Ty)) 2297 CO |= ClassOptions::HasConstructorOrDestructor; 2298 2299 std::string FullName = getFullyQualifiedName(Ty); 2300 2301 uint64_t SizeInBytes = Ty->getSizeInBits() / 8; 2302 2303 ClassRecord CR(Kind, FieldCount, CO, FieldTI, TypeIndex(), VShapeTI, 2304 SizeInBytes, FullName, Ty->getIdentifier()); 2305 TypeIndex ClassTI = TypeTable.writeLeafType(CR); 2306 2307 addUDTSrcLine(Ty, ClassTI); 2308 2309 addToUDTs(Ty); 2310 2311 return ClassTI; 2312 } 2313 2314 TypeIndex CodeViewDebug::lowerTypeUnion(const DICompositeType *Ty) { 2315 // Emit the complete type for unnamed unions. 2316 if (shouldAlwaysEmitCompleteClassType(Ty)) 2317 return getCompleteTypeIndex(Ty); 2318 2319 ClassOptions CO = 2320 ClassOptions::ForwardReference | getCommonClassOptions(Ty); 2321 std::string FullName = getFullyQualifiedName(Ty); 2322 UnionRecord UR(0, CO, TypeIndex(), 0, FullName, Ty->getIdentifier()); 2323 TypeIndex FwdDeclTI = TypeTable.writeLeafType(UR); 2324 if (!Ty->isForwardDecl()) 2325 DeferredCompleteTypes.push_back(Ty); 2326 return FwdDeclTI; 2327 } 2328 2329 TypeIndex CodeViewDebug::lowerCompleteTypeUnion(const DICompositeType *Ty) { 2330 ClassOptions CO = ClassOptions::Sealed | getCommonClassOptions(Ty); 2331 TypeIndex FieldTI; 2332 unsigned FieldCount; 2333 bool ContainsNestedClass; 2334 std::tie(FieldTI, std::ignore, FieldCount, ContainsNestedClass) = 2335 lowerRecordFieldList(Ty); 2336 2337 if (ContainsNestedClass) 2338 CO |= ClassOptions::ContainsNestedClass; 2339 2340 uint64_t SizeInBytes = Ty->getSizeInBits() / 8; 2341 std::string FullName = getFullyQualifiedName(Ty); 2342 2343 UnionRecord UR(FieldCount, CO, FieldTI, SizeInBytes, FullName, 2344 Ty->getIdentifier()); 2345 TypeIndex UnionTI = TypeTable.writeLeafType(UR); 2346 2347 addUDTSrcLine(Ty, UnionTI); 2348 2349 addToUDTs(Ty); 2350 2351 return UnionTI; 2352 } 2353 2354 std::tuple<TypeIndex, TypeIndex, unsigned, bool> 2355 CodeViewDebug::lowerRecordFieldList(const DICompositeType *Ty) { 2356 // Manually count members. MSVC appears to count everything that generates a 2357 // field list record. Each individual overload in a method overload group 2358 // contributes to this count, even though the overload group is a single field 2359 // list record. 2360 unsigned MemberCount = 0; 2361 ClassInfo Info = collectClassInfo(Ty); 2362 ContinuationRecordBuilder ContinuationBuilder; 2363 ContinuationBuilder.begin(ContinuationRecordKind::FieldList); 2364 2365 // Create base classes. 2366 for (const DIDerivedType *I : Info.Inheritance) { 2367 if (I->getFlags() & DINode::FlagVirtual) { 2368 // Virtual base. 2369 unsigned VBPtrOffset = I->getVBPtrOffset(); 2370 // FIXME: Despite the accessor name, the offset is really in bytes. 2371 unsigned VBTableIndex = I->getOffsetInBits() / 4; 2372 auto RecordKind = (I->getFlags() & DINode::FlagIndirectVirtualBase) == DINode::FlagIndirectVirtualBase 2373 ? TypeRecordKind::IndirectVirtualBaseClass 2374 : TypeRecordKind::VirtualBaseClass; 2375 VirtualBaseClassRecord VBCR( 2376 RecordKind, translateAccessFlags(Ty->getTag(), I->getFlags()), 2377 getTypeIndex(I->getBaseType()), getVBPTypeIndex(), VBPtrOffset, 2378 VBTableIndex); 2379 2380 ContinuationBuilder.writeMemberType(VBCR); 2381 MemberCount++; 2382 } else { 2383 assert(I->getOffsetInBits() % 8 == 0 && 2384 "bases must be on byte boundaries"); 2385 BaseClassRecord BCR(translateAccessFlags(Ty->getTag(), I->getFlags()), 2386 getTypeIndex(I->getBaseType()), 2387 I->getOffsetInBits() / 8); 2388 ContinuationBuilder.writeMemberType(BCR); 2389 MemberCount++; 2390 } 2391 } 2392 2393 // Create members. 2394 for (ClassInfo::MemberInfo &MemberInfo : Info.Members) { 2395 const DIDerivedType *Member = MemberInfo.MemberTypeNode; 2396 TypeIndex MemberBaseType = getTypeIndex(Member->getBaseType()); 2397 StringRef MemberName = Member->getName(); 2398 MemberAccess Access = 2399 translateAccessFlags(Ty->getTag(), Member->getFlags()); 2400 2401 if (Member->isStaticMember()) { 2402 StaticDataMemberRecord SDMR(Access, MemberBaseType, MemberName); 2403 ContinuationBuilder.writeMemberType(SDMR); 2404 MemberCount++; 2405 continue; 2406 } 2407 2408 // Virtual function pointer member. 2409 if ((Member->getFlags() & DINode::FlagArtificial) && 2410 Member->getName().startswith("_vptr$")) { 2411 VFPtrRecord VFPR(getTypeIndex(Member->getBaseType())); 2412 ContinuationBuilder.writeMemberType(VFPR); 2413 MemberCount++; 2414 continue; 2415 } 2416 2417 // Data member. 2418 uint64_t MemberOffsetInBits = 2419 Member->getOffsetInBits() + MemberInfo.BaseOffset; 2420 if (Member->isBitField()) { 2421 uint64_t StartBitOffset = MemberOffsetInBits; 2422 if (const auto *CI = 2423 dyn_cast_or_null<ConstantInt>(Member->getStorageOffsetInBits())) { 2424 MemberOffsetInBits = CI->getZExtValue() + MemberInfo.BaseOffset; 2425 } 2426 StartBitOffset -= MemberOffsetInBits; 2427 BitFieldRecord BFR(MemberBaseType, Member->getSizeInBits(), 2428 StartBitOffset); 2429 MemberBaseType = TypeTable.writeLeafType(BFR); 2430 } 2431 uint64_t MemberOffsetInBytes = MemberOffsetInBits / 8; 2432 DataMemberRecord DMR(Access, MemberBaseType, MemberOffsetInBytes, 2433 MemberName); 2434 ContinuationBuilder.writeMemberType(DMR); 2435 MemberCount++; 2436 } 2437 2438 // Create methods 2439 for (auto &MethodItr : Info.Methods) { 2440 StringRef Name = MethodItr.first->getString(); 2441 2442 std::vector<OneMethodRecord> Methods; 2443 for (const DISubprogram *SP : MethodItr.second) { 2444 TypeIndex MethodType = getMemberFunctionType(SP, Ty); 2445 bool Introduced = SP->getFlags() & DINode::FlagIntroducedVirtual; 2446 2447 unsigned VFTableOffset = -1; 2448 if (Introduced) 2449 VFTableOffset = SP->getVirtualIndex() * getPointerSizeInBytes(); 2450 2451 Methods.push_back(OneMethodRecord( 2452 MethodType, translateAccessFlags(Ty->getTag(), SP->getFlags()), 2453 translateMethodKindFlags(SP, Introduced), 2454 translateMethodOptionFlags(SP), VFTableOffset, Name)); 2455 MemberCount++; 2456 } 2457 assert(!Methods.empty() && "Empty methods map entry"); 2458 if (Methods.size() == 1) 2459 ContinuationBuilder.writeMemberType(Methods[0]); 2460 else { 2461 // FIXME: Make this use its own ContinuationBuilder so that 2462 // MethodOverloadList can be split correctly. 2463 MethodOverloadListRecord MOLR(Methods); 2464 TypeIndex MethodList = TypeTable.writeLeafType(MOLR); 2465 2466 OverloadedMethodRecord OMR(Methods.size(), MethodList, Name); 2467 ContinuationBuilder.writeMemberType(OMR); 2468 } 2469 } 2470 2471 // Create nested classes. 2472 for (const DIType *Nested : Info.NestedTypes) { 2473 NestedTypeRecord R(getTypeIndex(Nested), Nested->getName()); 2474 ContinuationBuilder.writeMemberType(R); 2475 MemberCount++; 2476 } 2477 2478 TypeIndex FieldTI = TypeTable.insertRecord(ContinuationBuilder); 2479 return std::make_tuple(FieldTI, Info.VShapeTI, MemberCount, 2480 !Info.NestedTypes.empty()); 2481 } 2482 2483 TypeIndex CodeViewDebug::getVBPTypeIndex() { 2484 if (!VBPType.getIndex()) { 2485 // Make a 'const int *' type. 2486 ModifierRecord MR(TypeIndex::Int32(), ModifierOptions::Const); 2487 TypeIndex ModifiedTI = TypeTable.writeLeafType(MR); 2488 2489 PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64 2490 : PointerKind::Near32; 2491 PointerMode PM = PointerMode::Pointer; 2492 PointerOptions PO = PointerOptions::None; 2493 PointerRecord PR(ModifiedTI, PK, PM, PO, getPointerSizeInBytes()); 2494 VBPType = TypeTable.writeLeafType(PR); 2495 } 2496 2497 return VBPType; 2498 } 2499 2500 TypeIndex CodeViewDebug::getTypeIndex(const DIType *Ty, const DIType *ClassTy) { 2501 // The null DIType is the void type. Don't try to hash it. 2502 if (!Ty) 2503 return TypeIndex::Void(); 2504 2505 // Check if we've already translated this type. Don't try to do a 2506 // get-or-create style insertion that caches the hash lookup across the 2507 // lowerType call. It will update the TypeIndices map. 2508 auto I = TypeIndices.find({Ty, ClassTy}); 2509 if (I != TypeIndices.end()) 2510 return I->second; 2511 2512 TypeLoweringScope S(*this); 2513 TypeIndex TI = lowerType(Ty, ClassTy); 2514 return recordTypeIndexForDINode(Ty, TI, ClassTy); 2515 } 2516 2517 codeview::TypeIndex 2518 CodeViewDebug::getTypeIndexForThisPtr(const DIDerivedType *PtrTy, 2519 const DISubroutineType *SubroutineTy) { 2520 assert(PtrTy->getTag() == dwarf::DW_TAG_pointer_type && 2521 "this type must be a pointer type"); 2522 2523 PointerOptions Options = PointerOptions::None; 2524 if (SubroutineTy->getFlags() & DINode::DIFlags::FlagLValueReference) 2525 Options = PointerOptions::LValueRefThisPointer; 2526 else if (SubroutineTy->getFlags() & DINode::DIFlags::FlagRValueReference) 2527 Options = PointerOptions::RValueRefThisPointer; 2528 2529 // Check if we've already translated this type. If there is no ref qualifier 2530 // on the function then we look up this pointer type with no associated class 2531 // so that the TypeIndex for the this pointer can be shared with the type 2532 // index for other pointers to this class type. If there is a ref qualifier 2533 // then we lookup the pointer using the subroutine as the parent type. 2534 auto I = TypeIndices.find({PtrTy, SubroutineTy}); 2535 if (I != TypeIndices.end()) 2536 return I->second; 2537 2538 TypeLoweringScope S(*this); 2539 TypeIndex TI = lowerTypePointer(PtrTy, Options); 2540 return recordTypeIndexForDINode(PtrTy, TI, SubroutineTy); 2541 } 2542 2543 TypeIndex CodeViewDebug::getTypeIndexForReferenceTo(const DIType *Ty) { 2544 PointerRecord PR(getTypeIndex(Ty), 2545 getPointerSizeInBytes() == 8 ? PointerKind::Near64 2546 : PointerKind::Near32, 2547 PointerMode::LValueReference, PointerOptions::None, 2548 Ty->getSizeInBits() / 8); 2549 return TypeTable.writeLeafType(PR); 2550 } 2551 2552 TypeIndex CodeViewDebug::getCompleteTypeIndex(const DIType *Ty) { 2553 // The null DIType is the void type. Don't try to hash it. 2554 if (!Ty) 2555 return TypeIndex::Void(); 2556 2557 // Look through typedefs when getting the complete type index. Call 2558 // getTypeIndex on the typdef to ensure that any UDTs are accumulated and are 2559 // emitted only once. 2560 if (Ty->getTag() == dwarf::DW_TAG_typedef) 2561 (void)getTypeIndex(Ty); 2562 while (Ty->getTag() == dwarf::DW_TAG_typedef) 2563 Ty = cast<DIDerivedType>(Ty)->getBaseType(); 2564 2565 // If this is a non-record type, the complete type index is the same as the 2566 // normal type index. Just call getTypeIndex. 2567 switch (Ty->getTag()) { 2568 case dwarf::DW_TAG_class_type: 2569 case dwarf::DW_TAG_structure_type: 2570 case dwarf::DW_TAG_union_type: 2571 break; 2572 default: 2573 return getTypeIndex(Ty); 2574 } 2575 2576 const auto *CTy = cast<DICompositeType>(Ty); 2577 2578 TypeLoweringScope S(*this); 2579 2580 // Make sure the forward declaration is emitted first. It's unclear if this 2581 // is necessary, but MSVC does it, and we should follow suit until we can show 2582 // otherwise. 2583 // We only emit a forward declaration for named types. 2584 if (!CTy->getName().empty() || !CTy->getIdentifier().empty()) { 2585 TypeIndex FwdDeclTI = getTypeIndex(CTy); 2586 2587 // Just use the forward decl if we don't have complete type info. This 2588 // might happen if the frontend is using modules and expects the complete 2589 // definition to be emitted elsewhere. 2590 if (CTy->isForwardDecl()) 2591 return FwdDeclTI; 2592 } 2593 2594 // Check if we've already translated the complete record type. 2595 // Insert the type with a null TypeIndex to signify that the type is currently 2596 // being lowered. 2597 auto InsertResult = CompleteTypeIndices.insert({CTy, TypeIndex()}); 2598 if (!InsertResult.second) 2599 return InsertResult.first->second; 2600 2601 TypeIndex TI; 2602 switch (CTy->getTag()) { 2603 case dwarf::DW_TAG_class_type: 2604 case dwarf::DW_TAG_structure_type: 2605 TI = lowerCompleteTypeClass(CTy); 2606 break; 2607 case dwarf::DW_TAG_union_type: 2608 TI = lowerCompleteTypeUnion(CTy); 2609 break; 2610 default: 2611 llvm_unreachable("not a record"); 2612 } 2613 2614 // Update the type index associated with this CompositeType. This cannot 2615 // use the 'InsertResult' iterator above because it is potentially 2616 // invalidated by map insertions which can occur while lowering the class 2617 // type above. 2618 CompleteTypeIndices[CTy] = TI; 2619 return TI; 2620 } 2621 2622 /// Emit all the deferred complete record types. Try to do this in FIFO order, 2623 /// and do this until fixpoint, as each complete record type typically 2624 /// references 2625 /// many other record types. 2626 void CodeViewDebug::emitDeferredCompleteTypes() { 2627 SmallVector<const DICompositeType *, 4> TypesToEmit; 2628 while (!DeferredCompleteTypes.empty()) { 2629 std::swap(DeferredCompleteTypes, TypesToEmit); 2630 for (const DICompositeType *RecordTy : TypesToEmit) 2631 getCompleteTypeIndex(RecordTy); 2632 TypesToEmit.clear(); 2633 } 2634 } 2635 2636 void CodeViewDebug::emitLocalVariableList(const FunctionInfo &FI, 2637 ArrayRef<LocalVariable> Locals) { 2638 // Get the sorted list of parameters and emit them first. 2639 SmallVector<const LocalVariable *, 6> Params; 2640 for (const LocalVariable &L : Locals) 2641 if (L.DIVar->isParameter()) 2642 Params.push_back(&L); 2643 llvm::sort(Params, [](const LocalVariable *L, const LocalVariable *R) { 2644 return L->DIVar->getArg() < R->DIVar->getArg(); 2645 }); 2646 for (const LocalVariable *L : Params) 2647 emitLocalVariable(FI, *L); 2648 2649 // Next emit all non-parameters in the order that we found them. 2650 for (const LocalVariable &L : Locals) 2651 if (!L.DIVar->isParameter()) 2652 emitLocalVariable(FI, L); 2653 } 2654 2655 void CodeViewDebug::emitLocalVariable(const FunctionInfo &FI, 2656 const LocalVariable &Var) { 2657 // LocalSym record, see SymbolRecord.h for more info. 2658 MCSymbol *LocalEnd = beginSymbolRecord(SymbolKind::S_LOCAL); 2659 2660 LocalSymFlags Flags = LocalSymFlags::None; 2661 if (Var.DIVar->isParameter()) 2662 Flags |= LocalSymFlags::IsParameter; 2663 if (Var.DefRanges.empty()) 2664 Flags |= LocalSymFlags::IsOptimizedOut; 2665 2666 OS.AddComment("TypeIndex"); 2667 TypeIndex TI = Var.UseReferenceType 2668 ? getTypeIndexForReferenceTo(Var.DIVar->getType()) 2669 : getCompleteTypeIndex(Var.DIVar->getType()); 2670 OS.emitInt32(TI.getIndex()); 2671 OS.AddComment("Flags"); 2672 OS.emitInt16(static_cast<uint16_t>(Flags)); 2673 // Truncate the name so we won't overflow the record length field. 2674 emitNullTerminatedSymbolName(OS, Var.DIVar->getName()); 2675 endSymbolRecord(LocalEnd); 2676 2677 // Calculate the on disk prefix of the appropriate def range record. The 2678 // records and on disk formats are described in SymbolRecords.h. BytePrefix 2679 // should be big enough to hold all forms without memory allocation. 2680 SmallString<20> BytePrefix; 2681 for (const LocalVarDefRange &DefRange : Var.DefRanges) { 2682 BytePrefix.clear(); 2683 if (DefRange.InMemory) { 2684 int Offset = DefRange.DataOffset; 2685 unsigned Reg = DefRange.CVRegister; 2686 2687 // 32-bit x86 call sequences often use PUSH instructions, which disrupt 2688 // ESP-relative offsets. Use the virtual frame pointer, VFRAME or $T0, 2689 // instead. In frames without stack realignment, $T0 will be the CFA. 2690 if (RegisterId(Reg) == RegisterId::ESP) { 2691 Reg = unsigned(RegisterId::VFRAME); 2692 Offset += FI.OffsetAdjustment; 2693 } 2694 2695 // If we can use the chosen frame pointer for the frame and this isn't a 2696 // sliced aggregate, use the smaller S_DEFRANGE_FRAMEPOINTER_REL record. 2697 // Otherwise, use S_DEFRANGE_REGISTER_REL. 2698 EncodedFramePtrReg EncFP = encodeFramePtrReg(RegisterId(Reg), TheCPU); 2699 if (!DefRange.IsSubfield && EncFP != EncodedFramePtrReg::None && 2700 (bool(Flags & LocalSymFlags::IsParameter) 2701 ? (EncFP == FI.EncodedParamFramePtrReg) 2702 : (EncFP == FI.EncodedLocalFramePtrReg))) { 2703 DefRangeFramePointerRelHeader DRHdr; 2704 DRHdr.Offset = Offset; 2705 OS.emitCVDefRangeDirective(DefRange.Ranges, DRHdr); 2706 } else { 2707 uint16_t RegRelFlags = 0; 2708 if (DefRange.IsSubfield) { 2709 RegRelFlags = DefRangeRegisterRelSym::IsSubfieldFlag | 2710 (DefRange.StructOffset 2711 << DefRangeRegisterRelSym::OffsetInParentShift); 2712 } 2713 DefRangeRegisterRelHeader DRHdr; 2714 DRHdr.Register = Reg; 2715 DRHdr.Flags = RegRelFlags; 2716 DRHdr.BasePointerOffset = Offset; 2717 OS.emitCVDefRangeDirective(DefRange.Ranges, DRHdr); 2718 } 2719 } else { 2720 assert(DefRange.DataOffset == 0 && "unexpected offset into register"); 2721 if (DefRange.IsSubfield) { 2722 DefRangeSubfieldRegisterHeader DRHdr; 2723 DRHdr.Register = DefRange.CVRegister; 2724 DRHdr.MayHaveNoName = 0; 2725 DRHdr.OffsetInParent = DefRange.StructOffset; 2726 OS.emitCVDefRangeDirective(DefRange.Ranges, DRHdr); 2727 } else { 2728 DefRangeRegisterHeader DRHdr; 2729 DRHdr.Register = DefRange.CVRegister; 2730 DRHdr.MayHaveNoName = 0; 2731 OS.emitCVDefRangeDirective(DefRange.Ranges, DRHdr); 2732 } 2733 } 2734 } 2735 } 2736 2737 void CodeViewDebug::emitLexicalBlockList(ArrayRef<LexicalBlock *> Blocks, 2738 const FunctionInfo& FI) { 2739 for (LexicalBlock *Block : Blocks) 2740 emitLexicalBlock(*Block, FI); 2741 } 2742 2743 /// Emit an S_BLOCK32 and S_END record pair delimiting the contents of a 2744 /// lexical block scope. 2745 void CodeViewDebug::emitLexicalBlock(const LexicalBlock &Block, 2746 const FunctionInfo& FI) { 2747 MCSymbol *RecordEnd = beginSymbolRecord(SymbolKind::S_BLOCK32); 2748 OS.AddComment("PtrParent"); 2749 OS.emitInt32(0); // PtrParent 2750 OS.AddComment("PtrEnd"); 2751 OS.emitInt32(0); // PtrEnd 2752 OS.AddComment("Code size"); 2753 OS.emitAbsoluteSymbolDiff(Block.End, Block.Begin, 4); // Code Size 2754 OS.AddComment("Function section relative address"); 2755 OS.EmitCOFFSecRel32(Block.Begin, /*Offset=*/0); // Func Offset 2756 OS.AddComment("Function section index"); 2757 OS.EmitCOFFSectionIndex(FI.Begin); // Func Symbol 2758 OS.AddComment("Lexical block name"); 2759 emitNullTerminatedSymbolName(OS, Block.Name); // Name 2760 endSymbolRecord(RecordEnd); 2761 2762 // Emit variables local to this lexical block. 2763 emitLocalVariableList(FI, Block.Locals); 2764 emitGlobalVariableList(Block.Globals); 2765 2766 // Emit lexical blocks contained within this block. 2767 emitLexicalBlockList(Block.Children, FI); 2768 2769 // Close the lexical block scope. 2770 emitEndSymbolRecord(SymbolKind::S_END); 2771 } 2772 2773 /// Convenience routine for collecting lexical block information for a list 2774 /// of lexical scopes. 2775 void CodeViewDebug::collectLexicalBlockInfo( 2776 SmallVectorImpl<LexicalScope *> &Scopes, 2777 SmallVectorImpl<LexicalBlock *> &Blocks, 2778 SmallVectorImpl<LocalVariable> &Locals, 2779 SmallVectorImpl<CVGlobalVariable> &Globals) { 2780 for (LexicalScope *Scope : Scopes) 2781 collectLexicalBlockInfo(*Scope, Blocks, Locals, Globals); 2782 } 2783 2784 /// Populate the lexical blocks and local variable lists of the parent with 2785 /// information about the specified lexical scope. 2786 void CodeViewDebug::collectLexicalBlockInfo( 2787 LexicalScope &Scope, 2788 SmallVectorImpl<LexicalBlock *> &ParentBlocks, 2789 SmallVectorImpl<LocalVariable> &ParentLocals, 2790 SmallVectorImpl<CVGlobalVariable> &ParentGlobals) { 2791 if (Scope.isAbstractScope()) 2792 return; 2793 2794 // Gather information about the lexical scope including local variables, 2795 // global variables, and address ranges. 2796 bool IgnoreScope = false; 2797 auto LI = ScopeVariables.find(&Scope); 2798 SmallVectorImpl<LocalVariable> *Locals = 2799 LI != ScopeVariables.end() ? &LI->second : nullptr; 2800 auto GI = ScopeGlobals.find(Scope.getScopeNode()); 2801 SmallVectorImpl<CVGlobalVariable> *Globals = 2802 GI != ScopeGlobals.end() ? GI->second.get() : nullptr; 2803 const DILexicalBlock *DILB = dyn_cast<DILexicalBlock>(Scope.getScopeNode()); 2804 const SmallVectorImpl<InsnRange> &Ranges = Scope.getRanges(); 2805 2806 // Ignore lexical scopes which do not contain variables. 2807 if (!Locals && !Globals) 2808 IgnoreScope = true; 2809 2810 // Ignore lexical scopes which are not lexical blocks. 2811 if (!DILB) 2812 IgnoreScope = true; 2813 2814 // Ignore scopes which have too many address ranges to represent in the 2815 // current CodeView format or do not have a valid address range. 2816 // 2817 // For lexical scopes with multiple address ranges you may be tempted to 2818 // construct a single range covering every instruction where the block is 2819 // live and everything in between. Unfortunately, Visual Studio only 2820 // displays variables from the first matching lexical block scope. If the 2821 // first lexical block contains exception handling code or cold code which 2822 // is moved to the bottom of the routine creating a single range covering 2823 // nearly the entire routine, then it will hide all other lexical blocks 2824 // and the variables they contain. 2825 if (Ranges.size() != 1 || !getLabelAfterInsn(Ranges.front().second)) 2826 IgnoreScope = true; 2827 2828 if (IgnoreScope) { 2829 // This scope can be safely ignored and eliminating it will reduce the 2830 // size of the debug information. Be sure to collect any variable and scope 2831 // information from the this scope or any of its children and collapse them 2832 // into the parent scope. 2833 if (Locals) 2834 ParentLocals.append(Locals->begin(), Locals->end()); 2835 if (Globals) 2836 ParentGlobals.append(Globals->begin(), Globals->end()); 2837 collectLexicalBlockInfo(Scope.getChildren(), 2838 ParentBlocks, 2839 ParentLocals, 2840 ParentGlobals); 2841 return; 2842 } 2843 2844 // Create a new CodeView lexical block for this lexical scope. If we've 2845 // seen this DILexicalBlock before then the scope tree is malformed and 2846 // we can handle this gracefully by not processing it a second time. 2847 auto BlockInsertion = CurFn->LexicalBlocks.insert({DILB, LexicalBlock()}); 2848 if (!BlockInsertion.second) 2849 return; 2850 2851 // Create a lexical block containing the variables and collect the the 2852 // lexical block information for the children. 2853 const InsnRange &Range = Ranges.front(); 2854 assert(Range.first && Range.second); 2855 LexicalBlock &Block = BlockInsertion.first->second; 2856 Block.Begin = getLabelBeforeInsn(Range.first); 2857 Block.End = getLabelAfterInsn(Range.second); 2858 assert(Block.Begin && "missing label for scope begin"); 2859 assert(Block.End && "missing label for scope end"); 2860 Block.Name = DILB->getName(); 2861 if (Locals) 2862 Block.Locals = std::move(*Locals); 2863 if (Globals) 2864 Block.Globals = std::move(*Globals); 2865 ParentBlocks.push_back(&Block); 2866 collectLexicalBlockInfo(Scope.getChildren(), 2867 Block.Children, 2868 Block.Locals, 2869 Block.Globals); 2870 } 2871 2872 void CodeViewDebug::endFunctionImpl(const MachineFunction *MF) { 2873 const Function &GV = MF->getFunction(); 2874 assert(FnDebugInfo.count(&GV)); 2875 assert(CurFn == FnDebugInfo[&GV].get()); 2876 2877 collectVariableInfo(GV.getSubprogram()); 2878 2879 // Build the lexical block structure to emit for this routine. 2880 if (LexicalScope *CFS = LScopes.getCurrentFunctionScope()) 2881 collectLexicalBlockInfo(*CFS, 2882 CurFn->ChildBlocks, 2883 CurFn->Locals, 2884 CurFn->Globals); 2885 2886 // Clear the scope and variable information from the map which will not be 2887 // valid after we have finished processing this routine. This also prepares 2888 // the map for the subsequent routine. 2889 ScopeVariables.clear(); 2890 2891 // Don't emit anything if we don't have any line tables. 2892 // Thunks are compiler-generated and probably won't have source correlation. 2893 if (!CurFn->HaveLineInfo && !GV.getSubprogram()->isThunk()) { 2894 FnDebugInfo.erase(&GV); 2895 CurFn = nullptr; 2896 return; 2897 } 2898 2899 // Find heap alloc sites and add to list. 2900 for (const auto &MBB : *MF) { 2901 for (const auto &MI : MBB) { 2902 if (MDNode *MD = MI.getHeapAllocMarker()) { 2903 CurFn->HeapAllocSites.push_back(std::make_tuple(getLabelBeforeInsn(&MI), 2904 getLabelAfterInsn(&MI), 2905 dyn_cast<DIType>(MD))); 2906 } 2907 } 2908 } 2909 2910 CurFn->Annotations = MF->getCodeViewAnnotations(); 2911 2912 CurFn->End = Asm->getFunctionEnd(); 2913 2914 CurFn = nullptr; 2915 } 2916 2917 // Usable locations are valid with non-zero line numbers. A line number of zero 2918 // corresponds to optimized code that doesn't have a distinct source location. 2919 // In this case, we try to use the previous or next source location depending on 2920 // the context. 2921 static bool isUsableDebugLoc(DebugLoc DL) { 2922 return DL && DL.getLine() != 0; 2923 } 2924 2925 void CodeViewDebug::beginInstruction(const MachineInstr *MI) { 2926 DebugHandlerBase::beginInstruction(MI); 2927 2928 // Ignore DBG_VALUE and DBG_LABEL locations and function prologue. 2929 if (!Asm || !CurFn || MI->isDebugInstr() || 2930 MI->getFlag(MachineInstr::FrameSetup)) 2931 return; 2932 2933 // If the first instruction of a new MBB has no location, find the first 2934 // instruction with a location and use that. 2935 DebugLoc DL = MI->getDebugLoc(); 2936 if (!isUsableDebugLoc(DL) && MI->getParent() != PrevInstBB) { 2937 for (const auto &NextMI : *MI->getParent()) { 2938 if (NextMI.isDebugInstr()) 2939 continue; 2940 DL = NextMI.getDebugLoc(); 2941 if (isUsableDebugLoc(DL)) 2942 break; 2943 } 2944 // FIXME: Handle the case where the BB has no valid locations. This would 2945 // probably require doing a real dataflow analysis. 2946 } 2947 PrevInstBB = MI->getParent(); 2948 2949 // If we still don't have a debug location, don't record a location. 2950 if (!isUsableDebugLoc(DL)) 2951 return; 2952 2953 maybeRecordLocation(DL, Asm->MF); 2954 } 2955 2956 MCSymbol *CodeViewDebug::beginCVSubsection(DebugSubsectionKind Kind) { 2957 MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(), 2958 *EndLabel = MMI->getContext().createTempSymbol(); 2959 OS.emitInt32(unsigned(Kind)); 2960 OS.AddComment("Subsection size"); 2961 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 4); 2962 OS.emitLabel(BeginLabel); 2963 return EndLabel; 2964 } 2965 2966 void CodeViewDebug::endCVSubsection(MCSymbol *EndLabel) { 2967 OS.emitLabel(EndLabel); 2968 // Every subsection must be aligned to a 4-byte boundary. 2969 OS.emitValueToAlignment(4); 2970 } 2971 2972 static StringRef getSymbolName(SymbolKind SymKind) { 2973 for (const EnumEntry<SymbolKind> &EE : getSymbolTypeNames()) 2974 if (EE.Value == SymKind) 2975 return EE.Name; 2976 return ""; 2977 } 2978 2979 MCSymbol *CodeViewDebug::beginSymbolRecord(SymbolKind SymKind) { 2980 MCSymbol *BeginLabel = MMI->getContext().createTempSymbol(), 2981 *EndLabel = MMI->getContext().createTempSymbol(); 2982 OS.AddComment("Record length"); 2983 OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2); 2984 OS.emitLabel(BeginLabel); 2985 if (OS.isVerboseAsm()) 2986 OS.AddComment("Record kind: " + getSymbolName(SymKind)); 2987 OS.emitInt16(unsigned(SymKind)); 2988 return EndLabel; 2989 } 2990 2991 void CodeViewDebug::endSymbolRecord(MCSymbol *SymEnd) { 2992 // MSVC does not pad out symbol records to four bytes, but LLVM does to avoid 2993 // an extra copy of every symbol record in LLD. This increases object file 2994 // size by less than 1% in the clang build, and is compatible with the Visual 2995 // C++ linker. 2996 OS.emitValueToAlignment(4); 2997 OS.emitLabel(SymEnd); 2998 } 2999 3000 void CodeViewDebug::emitEndSymbolRecord(SymbolKind EndKind) { 3001 OS.AddComment("Record length"); 3002 OS.emitInt16(2); 3003 if (OS.isVerboseAsm()) 3004 OS.AddComment("Record kind: " + getSymbolName(EndKind)); 3005 OS.emitInt16(uint16_t(EndKind)); // Record Kind 3006 } 3007 3008 void CodeViewDebug::emitDebugInfoForUDTs( 3009 const std::vector<std::pair<std::string, const DIType *>> &UDTs) { 3010 #ifndef NDEBUG 3011 size_t OriginalSize = UDTs.size(); 3012 #endif 3013 for (const auto &UDT : UDTs) { 3014 const DIType *T = UDT.second; 3015 assert(shouldEmitUdt(T)); 3016 MCSymbol *UDTRecordEnd = beginSymbolRecord(SymbolKind::S_UDT); 3017 OS.AddComment("Type"); 3018 OS.emitInt32(getCompleteTypeIndex(T).getIndex()); 3019 assert(OriginalSize == UDTs.size() && 3020 "getCompleteTypeIndex found new UDTs!"); 3021 emitNullTerminatedSymbolName(OS, UDT.first); 3022 endSymbolRecord(UDTRecordEnd); 3023 } 3024 } 3025 3026 void CodeViewDebug::collectGlobalVariableInfo() { 3027 DenseMap<const DIGlobalVariableExpression *, const GlobalVariable *> 3028 GlobalMap; 3029 for (const GlobalVariable &GV : MMI->getModule()->globals()) { 3030 SmallVector<DIGlobalVariableExpression *, 1> GVEs; 3031 GV.getDebugInfo(GVEs); 3032 for (const auto *GVE : GVEs) 3033 GlobalMap[GVE] = &GV; 3034 } 3035 3036 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 3037 for (const MDNode *Node : CUs->operands()) { 3038 const auto *CU = cast<DICompileUnit>(Node); 3039 for (const auto *GVE : CU->getGlobalVariables()) { 3040 const DIGlobalVariable *DIGV = GVE->getVariable(); 3041 const DIExpression *DIE = GVE->getExpression(); 3042 3043 // Emit constant global variables in a global symbol section. 3044 if (GlobalMap.count(GVE) == 0 && DIE->isConstant()) { 3045 CVGlobalVariable CVGV = {DIGV, DIE}; 3046 GlobalVariables.emplace_back(std::move(CVGV)); 3047 } 3048 3049 const auto *GV = GlobalMap.lookup(GVE); 3050 if (!GV || GV->isDeclarationForLinker()) 3051 continue; 3052 3053 DIScope *Scope = DIGV->getScope(); 3054 SmallVector<CVGlobalVariable, 1> *VariableList; 3055 if (Scope && isa<DILocalScope>(Scope)) { 3056 // Locate a global variable list for this scope, creating one if 3057 // necessary. 3058 auto Insertion = ScopeGlobals.insert( 3059 {Scope, std::unique_ptr<GlobalVariableList>()}); 3060 if (Insertion.second) 3061 Insertion.first->second = std::make_unique<GlobalVariableList>(); 3062 VariableList = Insertion.first->second.get(); 3063 } else if (GV->hasComdat()) 3064 // Emit this global variable into a COMDAT section. 3065 VariableList = &ComdatVariables; 3066 else 3067 // Emit this global variable in a single global symbol section. 3068 VariableList = &GlobalVariables; 3069 CVGlobalVariable CVGV = {DIGV, GV}; 3070 VariableList->emplace_back(std::move(CVGV)); 3071 } 3072 } 3073 } 3074 3075 void CodeViewDebug::emitDebugInfoForGlobals() { 3076 // First, emit all globals that are not in a comdat in a single symbol 3077 // substream. MSVC doesn't like it if the substream is empty, so only open 3078 // it if we have at least one global to emit. 3079 switchToDebugSectionForSymbol(nullptr); 3080 if (!GlobalVariables.empty()) { 3081 OS.AddComment("Symbol subsection for globals"); 3082 MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols); 3083 emitGlobalVariableList(GlobalVariables); 3084 endCVSubsection(EndLabel); 3085 } 3086 3087 // Second, emit each global that is in a comdat into its own .debug$S 3088 // section along with its own symbol substream. 3089 for (const CVGlobalVariable &CVGV : ComdatVariables) { 3090 const GlobalVariable *GV = CVGV.GVInfo.get<const GlobalVariable *>(); 3091 MCSymbol *GVSym = Asm->getSymbol(GV); 3092 OS.AddComment("Symbol subsection for " + 3093 Twine(GlobalValue::dropLLVMManglingEscape(GV->getName()))); 3094 switchToDebugSectionForSymbol(GVSym); 3095 MCSymbol *EndLabel = beginCVSubsection(DebugSubsectionKind::Symbols); 3096 // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. 3097 emitDebugInfoForGlobal(CVGV); 3098 endCVSubsection(EndLabel); 3099 } 3100 } 3101 3102 void CodeViewDebug::emitDebugInfoForRetainedTypes() { 3103 NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("llvm.dbg.cu"); 3104 for (const MDNode *Node : CUs->operands()) { 3105 for (auto *Ty : cast<DICompileUnit>(Node)->getRetainedTypes()) { 3106 if (DIType *RT = dyn_cast<DIType>(Ty)) { 3107 getTypeIndex(RT); 3108 // FIXME: Add to global/local DTU list. 3109 } 3110 } 3111 } 3112 } 3113 3114 // Emit each global variable in the specified array. 3115 void CodeViewDebug::emitGlobalVariableList(ArrayRef<CVGlobalVariable> Globals) { 3116 for (const CVGlobalVariable &CVGV : Globals) { 3117 // FIXME: emitDebugInfoForGlobal() doesn't handle DIExpressions. 3118 emitDebugInfoForGlobal(CVGV); 3119 } 3120 } 3121 3122 void CodeViewDebug::emitDebugInfoForGlobal(const CVGlobalVariable &CVGV) { 3123 const DIGlobalVariable *DIGV = CVGV.DIGV; 3124 3125 const DIScope *Scope = DIGV->getScope(); 3126 // For static data members, get the scope from the declaration. 3127 if (const auto *MemberDecl = dyn_cast_or_null<DIDerivedType>( 3128 DIGV->getRawStaticDataMemberDeclaration())) 3129 Scope = MemberDecl->getScope(); 3130 std::string QualifiedName = getFullyQualifiedName(Scope, DIGV->getName()); 3131 3132 if (const GlobalVariable *GV = 3133 CVGV.GVInfo.dyn_cast<const GlobalVariable *>()) { 3134 // DataSym record, see SymbolRecord.h for more info. Thread local data 3135 // happens to have the same format as global data. 3136 MCSymbol *GVSym = Asm->getSymbol(GV); 3137 SymbolKind DataSym = GV->isThreadLocal() 3138 ? (DIGV->isLocalToUnit() ? SymbolKind::S_LTHREAD32 3139 : SymbolKind::S_GTHREAD32) 3140 : (DIGV->isLocalToUnit() ? SymbolKind::S_LDATA32 3141 : SymbolKind::S_GDATA32); 3142 MCSymbol *DataEnd = beginSymbolRecord(DataSym); 3143 OS.AddComment("Type"); 3144 OS.emitInt32(getCompleteTypeIndex(DIGV->getType()).getIndex()); 3145 OS.AddComment("DataOffset"); 3146 OS.EmitCOFFSecRel32(GVSym, /*Offset=*/0); 3147 OS.AddComment("Segment"); 3148 OS.EmitCOFFSectionIndex(GVSym); 3149 OS.AddComment("Name"); 3150 const unsigned LengthOfDataRecord = 12; 3151 emitNullTerminatedSymbolName(OS, QualifiedName, LengthOfDataRecord); 3152 endSymbolRecord(DataEnd); 3153 } else { 3154 const DIExpression *DIE = CVGV.GVInfo.get<const DIExpression *>(); 3155 assert(DIE->isConstant() && 3156 "Global constant variables must contain a constant expression."); 3157 uint64_t Val = DIE->getElement(1); 3158 3159 MCSymbol *SConstantEnd = beginSymbolRecord(SymbolKind::S_CONSTANT); 3160 OS.AddComment("Type"); 3161 OS.emitInt32(getTypeIndex(DIGV->getType()).getIndex()); 3162 OS.AddComment("Value"); 3163 3164 // Encoded integers shouldn't need more than 10 bytes. 3165 uint8_t data[10]; 3166 BinaryStreamWriter Writer(data, llvm::support::endianness::little); 3167 CodeViewRecordIO IO(Writer); 3168 cantFail(IO.mapEncodedInteger(Val)); 3169 StringRef SRef((char *)data, Writer.getOffset()); 3170 OS.emitBinaryData(SRef); 3171 3172 OS.AddComment("Name"); 3173 emitNullTerminatedSymbolName(OS, QualifiedName); 3174 endSymbolRecord(SConstantEnd); 3175 } 3176 } 3177