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