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