blob: d77864c0466a05cbe00017e77ee3a2d440c4a5cb [file] [log] [blame]
//===- llvm/lib/CodeGen/AsmPrinter/CodeViewDebug.cpp ----------------------===//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
// This file contains support for writing Microsoft CodeView debug info.
#include "CodeViewDebug.h"
#include "DwarfExpression.h"
#include "llvm/ADT/APSInt.h"
#include "llvm/ADT/None.h"
#include "llvm/ADT/Optional.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/ADT/SmallString.h"
#include "llvm/ADT/StringRef.h"
#include "llvm/ADT/TinyPtrVector.h"
#include "llvm/ADT/Triple.h"
#include "llvm/ADT/Twine.h"
#include "llvm/BinaryFormat/COFF.h"
#include "llvm/BinaryFormat/Dwarf.h"
#include "llvm/CodeGen/AsmPrinter.h"
#include "llvm/CodeGen/LexicalScopes.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/MachineModuleInfo.h"
#include "llvm/CodeGen/MachineOperand.h"
#include "llvm/CodeGen/TargetFrameLowering.h"
#include "llvm/CodeGen/TargetRegisterInfo.h"
#include "llvm/CodeGen/TargetSubtargetInfo.h"
#include "llvm/Config/llvm-config.h"
#include "llvm/DebugInfo/CodeView/CVTypeVisitor.h"
#include "llvm/DebugInfo/CodeView/CodeViewRecordIO.h"
#include "llvm/DebugInfo/CodeView/ContinuationRecordBuilder.h"
#include "llvm/DebugInfo/CodeView/DebugInlineeLinesSubsection.h"
#include "llvm/DebugInfo/CodeView/EnumTables.h"
#include "llvm/DebugInfo/CodeView/Line.h"
#include "llvm/DebugInfo/CodeView/SymbolRecord.h"
#include "llvm/DebugInfo/CodeView/TypeDumpVisitor.h"
#include "llvm/DebugInfo/CodeView/TypeRecord.h"
#include "llvm/DebugInfo/CodeView/TypeTableCollection.h"
#include "llvm/DebugInfo/CodeView/TypeVisitorCallbackPipeline.h"
#include "llvm/IR/Constants.h"
#include "llvm/IR/DataLayout.h"
#include "llvm/IR/DebugInfoMetadata.h"
#include "llvm/IR/Function.h"
#include "llvm/IR/GlobalValue.h"
#include "llvm/IR/GlobalVariable.h"
#include "llvm/IR/Metadata.h"
#include "llvm/IR/Module.h"
#include "llvm/MC/MCAsmInfo.h"
#include "llvm/MC/MCContext.h"
#include "llvm/MC/MCSectionCOFF.h"
#include "llvm/MC/MCStreamer.h"
#include "llvm/MC/MCSymbol.h"
#include "llvm/Support/BinaryByteStream.h"
#include "llvm/Support/BinaryStreamReader.h"
#include "llvm/Support/BinaryStreamWriter.h"
#include "llvm/Support/Casting.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Endian.h"
#include "llvm/Support/Error.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/FormatVariadic.h"
#include "llvm/Support/Path.h"
#include "llvm/Support/SMLoc.h"
#include "llvm/Support/ScopedPrinter.h"
#include "llvm/Target/TargetLoweringObjectFile.h"
#include "llvm/Target/TargetMachine.h"
#include <algorithm>
#include <cassert>
#include <cctype>
#include <cstddef>
#include <iterator>
#include <limits>
using namespace llvm;
using namespace llvm::codeview;
namespace {
class CVMCAdapter : public CodeViewRecordStreamer {
CVMCAdapter(MCStreamer &OS, TypeCollection &TypeTable)
: OS(&OS), TypeTable(TypeTable) {}
void emitBytes(StringRef Data) override { OS->emitBytes(Data); }
void emitIntValue(uint64_t Value, unsigned Size) override {
OS->emitIntValueInHex(Value, Size);
void emitBinaryData(StringRef Data) override { OS->emitBinaryData(Data); }
void AddComment(const Twine &T) override { OS->AddComment(T); }
void AddRawComment(const Twine &T) override { OS->emitRawComment(T); }
bool isVerboseAsm() override { return OS->isVerboseAsm(); }
std::string getTypeName(TypeIndex TI) override {
std::string TypeName;
if (!TI.isNoneType()) {
if (TI.isSimple())
TypeName = std::string(TypeIndex::simpleTypeName(TI));
TypeName = std::string(TypeTable.getTypeName(TI));
return TypeName;
MCStreamer *OS = nullptr;
TypeCollection &TypeTable;
} // namespace
static CPUType mapArchToCVCPUType(Triple::ArchType Type) {
switch (Type) {
case Triple::ArchType::x86:
return CPUType::Pentium3;
case Triple::ArchType::x86_64:
return CPUType::X64;
case Triple::ArchType::thumb:
// LLVM currently doesn't support Windows CE and so thumb
// here is indiscriminately mapped to ARMNT specifically.
return CPUType::ARMNT;
case Triple::ArchType::aarch64:
return CPUType::ARM64;
report_fatal_error("target architecture doesn't map to a CodeView CPUType");
CodeViewDebug::CodeViewDebug(AsmPrinter *AP)
: DebugHandlerBase(AP), OS(*Asm->OutStreamer), TypeTable(Allocator) {}
StringRef CodeViewDebug::getFullFilepath(const DIFile *File) {
std::string &Filepath = FileToFilepathMap[File];
if (!Filepath.empty())
return Filepath;
StringRef Dir = File->getDirectory(), Filename = File->getFilename();
// If this is a Unix-style path, just use it as is. Don't try to canonicalize
// it textually because one of the path components could be a symlink.
if (Dir.startswith("/") || Filename.startswith("/")) {
if (llvm::sys::path::is_absolute(Filename, llvm::sys::path::Style::posix))
return Filename;
Filepath = std::string(Dir);
if (Dir.back() != '/')
Filepath += '/';
Filepath += Filename;
return Filepath;
// Clang emits directory and relative filename info into the IR, but CodeView
// operates on full paths. We could change Clang to emit full paths too, but
// that would increase the IR size and probably not needed for other users.
// For now, just concatenate and canonicalize the path here.
if (Filename.find(':') == 1)
Filepath = std::string(Filename);
Filepath = (Dir + "\\" + Filename).str();
// Canonicalize the path. We have to do it textually because we may no longer
// have access the file in the filesystem.
// First, replace all slashes with backslashes.
std::replace(Filepath.begin(), Filepath.end(), '/', '\\');
// Remove all "\.\" with "\".
size_t Cursor = 0;
while ((Cursor = Filepath.find("\\.\\", Cursor)) != std::string::npos)
Filepath.erase(Cursor, 2);
// Replace all "\XXX\..\" with "\". Don't try too hard though as the original
// path should be well-formatted, e.g. start with a drive letter, etc.
Cursor = 0;
while ((Cursor = Filepath.find("\\..\\", Cursor)) != std::string::npos) {
// Something's wrong if the path starts with "\..\", abort.
if (Cursor == 0)
size_t PrevSlash = Filepath.rfind('\\', Cursor - 1);
if (PrevSlash == std::string::npos)
// Something's wrong, abort.
Filepath.erase(PrevSlash, Cursor + 3 - PrevSlash);
// The next ".." might be following the one we've just erased.
Cursor = PrevSlash;
// Remove all duplicate backslashes.
Cursor = 0;
while ((Cursor = Filepath.find("\\\\", Cursor)) != std::string::npos)
Filepath.erase(Cursor, 1);
return Filepath;
unsigned CodeViewDebug::maybeRecordFile(const DIFile *F) {
StringRef FullPath = getFullFilepath(F);
unsigned NextId = FileIdMap.size() + 1;
auto Insertion = FileIdMap.insert(std::make_pair(FullPath, NextId));
if (Insertion.second) {
// We have to compute the full filepath and emit a .cv_file directive.
ArrayRef<uint8_t> ChecksumAsBytes;
FileChecksumKind CSKind = FileChecksumKind::None;
if (F->getChecksum()) {
std::string Checksum = fromHex(F->getChecksum()->Value);
void *CKMem = OS.getContext().allocate(Checksum.size(), 1);
memcpy(CKMem,, Checksum.size());
ChecksumAsBytes = ArrayRef<uint8_t>(
reinterpret_cast<const uint8_t *>(CKMem), Checksum.size());
switch (F->getChecksum()->Kind) {
case DIFile::CSK_MD5:
CSKind = FileChecksumKind::MD5;
case DIFile::CSK_SHA1:
CSKind = FileChecksumKind::SHA1;
case DIFile::CSK_SHA256:
CSKind = FileChecksumKind::SHA256;
bool Success = OS.EmitCVFileDirective(NextId, FullPath, ChecksumAsBytes,
assert(Success && ".cv_file directive failed");
return Insertion.first->second;
CodeViewDebug::InlineSite &
CodeViewDebug::getInlineSite(const DILocation *InlinedAt,
const DISubprogram *Inlinee) {
auto SiteInsertion = CurFn->InlineSites.insert({InlinedAt, InlineSite()});
InlineSite *Site = &SiteInsertion.first->second;
if (SiteInsertion.second) {
unsigned ParentFuncId = CurFn->FuncId;
if (const DILocation *OuterIA = InlinedAt->getInlinedAt())
ParentFuncId =
getInlineSite(OuterIA, InlinedAt->getScope()->getSubprogram())
Site->SiteFuncId = NextFuncId++;
Site->SiteFuncId, ParentFuncId, maybeRecordFile(InlinedAt->getFile()),
InlinedAt->getLine(), InlinedAt->getColumn(), SMLoc());
Site->Inlinee = Inlinee;
return *Site;
static StringRef getPrettyScopeName(const DIScope *Scope) {
StringRef ScopeName = Scope->getName();
if (!ScopeName.empty())
return ScopeName;
switch (Scope->getTag()) {
case dwarf::DW_TAG_enumeration_type:
case dwarf::DW_TAG_class_type:
case dwarf::DW_TAG_structure_type:
case dwarf::DW_TAG_union_type:
return "<unnamed-tag>";
case dwarf::DW_TAG_namespace:
return "`anonymous namespace'";
return StringRef();
const DISubprogram *CodeViewDebug::collectParentScopeNames(
const DIScope *Scope, SmallVectorImpl<StringRef> &QualifiedNameComponents) {
const DISubprogram *ClosestSubprogram = nullptr;
while (Scope != nullptr) {
if (ClosestSubprogram == nullptr)
ClosestSubprogram = dyn_cast<DISubprogram>(Scope);
// If a type appears in a scope chain, make sure it gets emitted. The
// frontend will be responsible for deciding if this should be a forward
// declaration or a complete type.
if (const auto *Ty = dyn_cast<DICompositeType>(Scope))
StringRef ScopeName = getPrettyScopeName(Scope);
if (!ScopeName.empty())
Scope = Scope->getScope();
return ClosestSubprogram;
static std::string formatNestedName(ArrayRef<StringRef> QualifiedNameComponents,
StringRef TypeName) {
std::string FullyQualifiedName;
for (StringRef QualifiedNameComponent :
llvm::reverse(QualifiedNameComponents)) {
return FullyQualifiedName;
struct CodeViewDebug::TypeLoweringScope {
TypeLoweringScope(CodeViewDebug &CVD) : CVD(CVD) { ++CVD.TypeEmissionLevel; }
~TypeLoweringScope() {
// Don't decrement TypeEmissionLevel until after emitting deferred types, so
// inner TypeLoweringScopes don't attempt to emit deferred types.
if (CVD.TypeEmissionLevel == 1)
CodeViewDebug &CVD;
std::string CodeViewDebug::getFullyQualifiedName(const DIScope *Scope,
StringRef Name) {
// Ensure types in the scope chain are emitted as soon as possible.
// This can create otherwise a situation where S_UDTs are emitted while
// looping in emitDebugInfoForUDTs.
TypeLoweringScope S(*this);
SmallVector<StringRef, 5> QualifiedNameComponents;
collectParentScopeNames(Scope, QualifiedNameComponents);
return formatNestedName(QualifiedNameComponents, Name);
std::string CodeViewDebug::getFullyQualifiedName(const DIScope *Ty) {
const DIScope *Scope = Ty->getScope();
return getFullyQualifiedName(Scope, getPrettyScopeName(Ty));
TypeIndex CodeViewDebug::getScopeIndex(const DIScope *Scope) {
// No scope means global scope and that uses the zero index.
if (!Scope || isa<DIFile>(Scope))
return TypeIndex();
assert(!isa<DIType>(Scope) && "shouldn't make a namespace scope for a type");
// Check if we've already translated this scope.
auto I = TypeIndices.find({Scope, nullptr});
if (I != TypeIndices.end())
return I->second;
// Build the fully qualified name of the scope.
std::string ScopeName = getFullyQualifiedName(Scope);
StringIdRecord SID(TypeIndex(), ScopeName);
auto TI = TypeTable.writeLeafType(SID);
return recordTypeIndexForDINode(Scope, TI);
static StringRef removeTemplateArgs(StringRef Name) {
// Remove template args from the display name. Assume that the template args
// are the last thing in the name.
if (Name.empty() || Name.back() != '>')
return Name;
int OpenBrackets = 0;
for (int i = Name.size() - 1; i >= 0; --i) {
if (Name[i] == '>')
else if (Name[i] == '<') {
if (OpenBrackets == 0)
return Name.substr(0, i);
return Name;
TypeIndex CodeViewDebug::getFuncIdForSubprogram(const DISubprogram *SP) {
// Check if we've already translated this subprogram.
auto I = TypeIndices.find({SP, nullptr});
if (I != TypeIndices.end())
return I->second;
// The display name includes function template arguments. Drop them to match
// MSVC. We need to have the template arguments in the DISubprogram name
// because they are used in other symbol records, such as S_GPROC32_IDs.
StringRef DisplayName = removeTemplateArgs(SP->getName());
const DIScope *Scope = SP->getScope();
TypeIndex TI;
if (const auto *Class = dyn_cast_or_null<DICompositeType>(Scope)) {
// If the scope is a DICompositeType, then this must be a method. Member
// function types take some special handling, and require access to the
// subprogram.
TypeIndex ClassType = getTypeIndex(Class);
MemberFuncIdRecord MFuncId(ClassType, getMemberFunctionType(SP, Class),
TI = TypeTable.writeLeafType(MFuncId);
} else {
// Otherwise, this must be a free function.
TypeIndex ParentScope = getScopeIndex(Scope);
FuncIdRecord FuncId(ParentScope, getTypeIndex(SP->getType()), DisplayName);
TI = TypeTable.writeLeafType(FuncId);
return recordTypeIndexForDINode(SP, TI);
static bool isNonTrivial(const DICompositeType *DCTy) {
return ((DCTy->getFlags() & DINode::FlagNonTrivial) == DINode::FlagNonTrivial);
static FunctionOptions
getFunctionOptions(const DISubroutineType *Ty,
const DICompositeType *ClassTy = nullptr,
StringRef SPName = StringRef("")) {
FunctionOptions FO = FunctionOptions::None;
const DIType *ReturnTy = nullptr;
if (auto TypeArray = Ty->getTypeArray()) {
if (TypeArray.size())
ReturnTy = TypeArray[0];
// Add CxxReturnUdt option to functions that return nontrivial record types
// or methods that return record types.
if (auto *ReturnDCTy = dyn_cast_or_null<DICompositeType>(ReturnTy))
if (isNonTrivial(ReturnDCTy) || ClassTy)
FO |= FunctionOptions::CxxReturnUdt;
// DISubroutineType is unnamed. Use DISubprogram's i.e. SPName in comparison.
if (ClassTy && isNonTrivial(ClassTy) && SPName == ClassTy->getName()) {
FO |= FunctionOptions::Constructor;
// TODO: put the FunctionOptions::ConstructorWithVirtualBases flag.
return FO;
TypeIndex CodeViewDebug::getMemberFunctionType(const DISubprogram *SP,
const DICompositeType *Class) {
// Always use the method declaration as the key for the function type. The
// method declaration contains the this adjustment.
if (SP->getDeclaration())
SP = SP->getDeclaration();
assert(!SP->getDeclaration() && "should use declaration as key");
// Key the MemberFunctionRecord into the map as {SP, Class}. It won't collide
// with the MemberFuncIdRecord, which is keyed in as {SP, nullptr}.
auto I = TypeIndices.find({SP, Class});
if (I != TypeIndices.end())
return I->second;
// Make sure complete type info for the class is emitted *after* the member
// function type, as the complete class type is likely to reference this
// member function type.
TypeLoweringScope S(*this);
const bool IsStaticMethod = (SP->getFlags() & DINode::FlagStaticMember) != 0;
FunctionOptions FO = getFunctionOptions(SP->getType(), Class, SP->getName());
TypeIndex TI = lowerTypeMemberFunction(
SP->getType(), Class, SP->getThisAdjustment(), IsStaticMethod, FO);
return recordTypeIndexForDINode(SP, TI, Class);
TypeIndex CodeViewDebug::recordTypeIndexForDINode(const DINode *Node,
TypeIndex TI,
const DIType *ClassTy) {
auto InsertResult = TypeIndices.insert({{Node, ClassTy}, TI});
assert(InsertResult.second && "DINode was already assigned a type index");
return TI;
unsigned CodeViewDebug::getPointerSizeInBytes() {
return MMI->getModule()->getDataLayout().getPointerSizeInBits() / 8;
void CodeViewDebug::recordLocalVariable(LocalVariable &&Var,
const LexicalScope *LS) {
if (const DILocation *InlinedAt = LS->getInlinedAt()) {
// This variable was inlined. Associate it with the InlineSite.
const DISubprogram *Inlinee = Var.DIVar->getScope()->getSubprogram();
InlineSite &Site = getInlineSite(InlinedAt, Inlinee);
} else {
// This variable goes into the corresponding lexical scope.
static void addLocIfNotPresent(SmallVectorImpl<const DILocation *> &Locs,
const DILocation *Loc) {
if (!llvm::is_contained(Locs, Loc))
void CodeViewDebug::maybeRecordLocation(const DebugLoc &DL,
const MachineFunction *MF) {
// Skip this instruction if it has the same location as the previous one.
if (!DL || DL == PrevInstLoc)
const DIScope *Scope = DL.get()->getScope();
if (!Scope)
// Skip this line if it is longer than the maximum we can record.
LineInfo LI(DL.getLine(), DL.getLine(), /*IsStatement=*/true);
if (LI.getStartLine() != DL.getLine() || LI.isAlwaysStepInto() ||
ColumnInfo CI(DL.getCol(), /*EndColumn=*/0);
if (CI.getStartColumn() != DL.getCol())
if (!CurFn->HaveLineInfo)
CurFn->HaveLineInfo = true;
unsigned FileId = 0;
if (PrevInstLoc.get() && PrevInstLoc->getFile() == DL->getFile())
FileId = CurFn->LastFileId;
FileId = CurFn->LastFileId = maybeRecordFile(DL->getFile());
PrevInstLoc = DL;
unsigned FuncId = CurFn->FuncId;
if (const DILocation *SiteLoc = DL->getInlinedAt()) {
const DILocation *Loc = DL.get();
// If this location was actually inlined from somewhere else, give it the ID
// of the inline call site.
FuncId =
getInlineSite(SiteLoc, Loc->getScope()->getSubprogram()).SiteFuncId;
// Ensure we have links in the tree of inline call sites.
bool FirstLoc = true;
while ((SiteLoc = Loc->getInlinedAt())) {
InlineSite &Site =
getInlineSite(SiteLoc, Loc->getScope()->getSubprogram());
if (!FirstLoc)
addLocIfNotPresent(Site.ChildSites, Loc);
FirstLoc = false;
Loc = SiteLoc;
addLocIfNotPresent(CurFn->ChildSites, Loc);
OS.emitCVLocDirective(FuncId, FileId, DL.getLine(), DL.getCol(),
/*PrologueEnd=*/false, /*IsStmt=*/false,
DL->getFilename(), SMLoc());
void CodeViewDebug::emitCodeViewMagicVersion() {
OS.AddComment("Debug section magic");
void CodeViewDebug::beginModule(Module *M) {
// If module doesn't have named metadata anchors or COFF debug section
// is not available, skip any debug info related stuff.
if (!M->getNamedMetadata("") ||
!Asm->getObjFileLowering().getCOFFDebugSymbolsSection()) {
Asm = nullptr;
// Tell MMI that we have and need debug info.
TheCPU = mapArchToCVCPUType(Triple(M->getTargetTriple()).getArch());
// Check if we should emit type record hashes.
ConstantInt *GH =
EmitDebugGlobalHashes = GH && !GH->isZero();
void CodeViewDebug::endModule() {
if (!Asm || !MMI->hasDebugInfo())
// The COFF .debug$S section consists of several subsections, each starting
// with a 4-byte control code (e.g. 0xF1, 0xF2, etc) and then a 4-byte length
// of the payload followed by the payload itself. The subsections are 4-byte
// aligned.
// Use the generic .debug$S section, and make a subsection for all the inlined
// subprograms.
MCSymbol *CompilerInfo = beginCVSubsection(DebugSubsectionKind::Symbols);
// Emit per-function debug information.
for (auto &P : FnDebugInfo)
if (!P.first->isDeclarationForLinker())
emitDebugInfoForFunction(P.first, *P.second);
// Get types used by globals without emitting anything.
// This is meant to collect all static const data members so they can be
// emitted as globals.
// Emit retained types.
// Emit global variable debug information.
// Switch back to the generic .debug$S section after potentially processing
// comdat symbol sections.
// Emit UDT records for any types used by global variables.
if (!GlobalUDTs.empty()) {
MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols);
// This subsection holds a file index to offset in string table table.
OS.AddComment("File index to string table offset subsection");
// This subsection holds the string table.
OS.AddComment("String table");
// Emit S_BUILDINFO, which points to LF_BUILDINFO. Put this in its own symbol
// subsection in the generic .debug$S section at the end. There is no
// particular reason for this ordering other than to match MSVC.
// Emit type information and hashes last, so that any types we translate while
// emitting function info are included.
if (EmitDebugGlobalHashes)
static void
emitNullTerminatedSymbolName(MCStreamer &OS, StringRef S,
unsigned MaxFixedRecordLength = 0xF00) {
// The maximum CV record length is 0xFF00. Most of the strings we emit appear
// after a fixed length portion of the record. The fixed length portion should
// always be less than 0xF00 (3840) bytes, so truncate the string so that the
// overall record size is less than the maximum allowed.
SmallString<32> NullTerminatedString(
S.take_front(MaxRecordLength - MaxFixedRecordLength - 1));
void CodeViewDebug::emitTypeInformation() {
if (TypeTable.empty())
// Start the .debug$T or .debug$P section with 0x4.
TypeTableCollection Table(TypeTable.records());
TypeVisitorCallbackPipeline Pipeline;
// To emit type record using Codeview MCStreamer adapter
CVMCAdapter CVMCOS(OS, Table);
TypeRecordMapping typeMapping(CVMCOS);
Optional<TypeIndex> B = Table.getFirst();
while (B) {
// This will fail if the record data is invalid.
CVType Record = Table.getType(*B);
Error E = codeview::visitTypeRecord(Record, *B, Pipeline);
if (E) {
logAllUnhandledErrors(std::move(E), errs(), "error: ");
llvm_unreachable("produced malformed type record");
B = Table.getNext(*B);
void CodeViewDebug::emitTypeGlobalHashes() {
if (TypeTable.empty())
// Start the .debug$H section with the version and hash algorithm, currently
// hardcoded to version 0, SHA1.
OS.AddComment("Section Version");
OS.AddComment("Hash Algorithm");
TypeIndex TI(TypeIndex::FirstNonSimpleIndex);
for (const auto &GHR : TypeTable.hashes()) {
if (OS.isVerboseAsm()) {
// Emit an EOL-comment describing which TypeIndex this hash corresponds
// to, as well as the stringified SHA1 hash.
SmallString<32> Comment;
raw_svector_ostream CommentOS(Comment);
CommentOS << formatv("{0:X+} [{1}]", TI.getIndex(), GHR);
assert(GHR.Hash.size() == 8);
StringRef S(reinterpret_cast<const char *>(,
static SourceLanguage MapDWLangToCVLang(unsigned DWLang) {
switch (DWLang) {
case dwarf::DW_LANG_C:
case dwarf::DW_LANG_C89:
case dwarf::DW_LANG_C99:
case dwarf::DW_LANG_C11:
case dwarf::DW_LANG_ObjC:
return SourceLanguage::C;
case dwarf::DW_LANG_C_plus_plus:
case dwarf::DW_LANG_C_plus_plus_03:
case dwarf::DW_LANG_C_plus_plus_11:
case dwarf::DW_LANG_C_plus_plus_14:
return SourceLanguage::Cpp;
case dwarf::DW_LANG_Fortran77:
case dwarf::DW_LANG_Fortran90:
case dwarf::DW_LANG_Fortran03:
case dwarf::DW_LANG_Fortran08:
return SourceLanguage::Fortran;
case dwarf::DW_LANG_Pascal83:
return SourceLanguage::Pascal;
case dwarf::DW_LANG_Cobol74:
case dwarf::DW_LANG_Cobol85:
return SourceLanguage::Cobol;
case dwarf::DW_LANG_Java:
return SourceLanguage::Java;
case dwarf::DW_LANG_D:
return SourceLanguage::D;
case dwarf::DW_LANG_Swift:
return SourceLanguage::Swift;
// There's no CodeView representation for this language, and CV doesn't
// have an "unknown" option for the language field, so we'll use MASM,
// as it's very low level.
return SourceLanguage::Masm;
namespace {
struct Version {
int Part[4];
} // end anonymous namespace
// Takes a StringRef like "clang (other nonsense 123)" and parses out
// the version number.
static Version parseVersion(StringRef Name) {
Version V = {{0}};
int N = 0;
for (const char C : Name) {
if (isdigit(C)) {
V.Part[N] *= 10;
V.Part[N] += C - '0';
} else if (C == '.') {
if (N >= 4)
return V;
} else if (N > 0)
return V;
return V;
void CodeViewDebug::emitCompilerInformation() {
MCSymbol *CompilerEnd = beginSymbolRecord(SymbolKind::S_COMPILE3);
uint32_t Flags = 0;
NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("");
const MDNode *Node = *CUs->operands().begin();
const auto *CU = cast<DICompileUnit>(Node);
// The low byte of the flags indicates the source language.
Flags = MapDWLangToCVLang(CU->getSourceLanguage());
// TODO: Figure out which other flags need to be set.
OS.AddComment("Flags and language");
StringRef CompilerVersion = CU->getProducer();
Version FrontVer = parseVersion(CompilerVersion);
OS.AddComment("Frontend version");
for (int N : FrontVer.Part)
// Some Microsoft tools, like Binscope, expect a backend version number of at
// least 8.something, so we'll coerce the LLVM version into a form that
// guarantees it'll be big enough without really lying about the version.
int Major = 1000 * LLVM_VERSION_MAJOR +
// Clamp it for builds that use unusually large version numbers.
Major = std::min<int>(Major, std::numeric_limits<uint16_t>::max());
Version BackVer = {{ Major, 0, 0, 0 }};
OS.AddComment("Backend version");
for (int N : BackVer.Part)
OS.AddComment("Null-terminated compiler version string");
emitNullTerminatedSymbolName(OS, CompilerVersion);
static TypeIndex getStringIdTypeIdx(GlobalTypeTableBuilder &TypeTable,
StringRef S) {
StringIdRecord SIR(TypeIndex(0x0), S);
return TypeTable.writeLeafType(SIR);
void CodeViewDebug::emitBuildInfo() {
// First, make LF_BUILDINFO. It's a sequence of strings with various bits of
// build info. The known prefix is:
// - Absolute path of current directory
// - Compiler path
// - Main source file path, relative to CWD or absolute
// - Type server PDB file
// - Canonical compiler command line
// If frontend and backend compilation are separated (think llc or LTO), it's
// not clear if the compiler path should refer to the executable for the
// frontend or the backend. Leave it blank for now.
TypeIndex BuildInfoArgs[BuildInfoRecord::MaxArgs] = {};
NamedMDNode *CUs = MMI->getModule()->getNamedMetadata("");
const MDNode *Node = *CUs->operands().begin(); // FIXME: Multiple CUs.
const auto *CU = cast<DICompileUnit>(Node);
const DIFile *MainSourceFile = CU->getFile();
BuildInfoArgs[BuildInfoRecord::CurrentDirectory] =
getStringIdTypeIdx(TypeTable, MainSourceFile->getDirectory());
BuildInfoArgs[BuildInfoRecord::SourceFile] =
getStringIdTypeIdx(TypeTable, MainSourceFile->getFilename());
// FIXME: Path to compiler and command line. PDB is intentionally blank unless
// we implement /Zi type servers.
BuildInfoRecord BIR(BuildInfoArgs);
TypeIndex BuildInfoIndex = TypeTable.writeLeafType(BIR);
// Make a new .debug$S subsection for the S_BUILDINFO record, which points
// from the module symbols into the type stream.
MCSymbol *BISubsecEnd = beginCVSubsection(DebugSubsectionKind::Symbols);
MCSymbol *BIEnd = beginSymbolRecord(SymbolKind::S_BUILDINFO);
OS.AddComment("LF_BUILDINFO index");
void CodeViewDebug::emitInlineeLinesSubsection() {
if (InlinedSubprograms.empty())
OS.AddComment("Inlinee lines subsection");
MCSymbol *InlineEnd = beginCVSubsection(DebugSubsectionKind::InlineeLines);
// We emit the checksum info for files. This is used by debuggers to
// determine if a pdb matches the source before loading it. Visual Studio,
// for instance, will display a warning that the breakpoints are not valid if
// the pdb does not match the source.
OS.AddComment("Inlinee lines signature");
for (const DISubprogram *SP : InlinedSubprograms) {
assert(TypeIndices.count({SP, nullptr}));
TypeIndex InlineeIdx = TypeIndices[{SP, nullptr}];
unsigned FileId = maybeRecordFile(SP->getFile());
OS.AddComment("Inlined function " + SP->getName() + " starts at " +
SP->getFilename() + Twine(':') + Twine(SP->getLine()));
OS.AddComment("Type index of inlined function");
OS.AddComment("Offset into filechecksum table");
OS.AddComment("Starting line number");
void CodeViewDebug::emitInlinedCallSite(const FunctionInfo &FI,
const DILocation *InlinedAt,
const InlineSite &Site) {
assert(TypeIndices.count({Site.Inlinee, nullptr}));
TypeIndex InlineeIdx = TypeIndices[{Site.Inlinee, nullptr}];
// SymbolRecord
MCSymbol *InlineEnd = beginSymbolRecord(SymbolKind::S_INLINESITE);
OS.AddComment("Inlinee type index");
unsigned FileId = maybeRecordFile(Site.Inlinee->getFile());
unsigned StartLineNum = Site.Inlinee->getLine();
OS.emitCVInlineLinetableDirective(Site.SiteFuncId, FileId, StartLineNum,
FI.Begin, FI.End);
emitLocalVariableList(FI, Site.InlinedLocals);
// Recurse on child inlined call sites before closing the scope.
for (const DILocation *ChildSite : Site.ChildSites) {
auto I = FI.InlineSites.find(ChildSite);
assert(I != FI.InlineSites.end() &&
"child site not in function inline site map");
emitInlinedCallSite(FI, ChildSite, I->second);
// Close the scope.
void CodeViewDebug::switchToDebugSectionForSymbol(const MCSymbol *GVSym) {
// If we have a symbol, it may be in a section that is COMDAT. If so, find the
// comdat key. A section may be comdat because of -ffunction-sections or
// because it is comdat in the IR.
MCSectionCOFF *GVSec =
GVSym ? dyn_cast<MCSectionCOFF>(&GVSym->getSection()) : nullptr;
const MCSymbol *KeySym = GVSec ? GVSec->getCOMDATSymbol() : nullptr;
MCSectionCOFF *DebugSec = cast<MCSectionCOFF>(
DebugSec = OS.getContext().getAssociativeCOFFSection(DebugSec, KeySym);
// Emit the magic version number if this is the first time we've switched to
// this section.
if (ComdatDebugSections.insert(DebugSec).second)
// Emit an S_THUNK32/S_END symbol pair for a thunk routine.
// The only supported thunk ordinal is currently the standard type.
void CodeViewDebug::emitDebugInfoForThunk(const Function *GV,
FunctionInfo &FI,
const MCSymbol *Fn) {
std::string FuncName =
const ThunkOrdinal ordinal = ThunkOrdinal::Standard; // Only supported kind.
OS.AddComment("Symbol subsection for " + Twine(FuncName));
MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols);
// Emit S_THUNK32
MCSymbol *ThunkRecordEnd = beginSymbolRecord(SymbolKind::S_THUNK32);
OS.AddComment("Thunk section relative address");
OS.EmitCOFFSecRel32(Fn, /*Offset=*/0);
OS.AddComment("Thunk section index");
OS.AddComment("Code size");
OS.emitAbsoluteSymbolDiff(FI.End, Fn, 2);
OS.AddComment("Function name");
emitNullTerminatedSymbolName(OS, FuncName);
// Additional fields specific to the thunk ordinal would go here.
// Local variables/inlined routines are purposely omitted here. The point of
// marking this as a thunk is so Visual Studio will NOT stop in this routine.
void CodeViewDebug::emitDebugInfoForFunction(const Function *GV,
FunctionInfo &FI) {
// For each function there is a separate subsection which holds the PC to
// file:line table.
const MCSymbol *Fn = Asm->getSymbol(GV);
// Switch to the to a comdat section, if appropriate.
std::string FuncName;
auto *SP = GV->getSubprogram();
if (SP->isThunk()) {
emitDebugInfoForThunk(GV, FI, Fn);
// If we have a display name, build the fully qualified name by walking the
// chain of scopes.
if (!SP->getName().empty())
FuncName = getFullyQualifiedName(SP->getScope(), SP->getName());
// If our DISubprogram name is empty, use the mangled name.
if (FuncName.empty())
FuncName = std::string(GlobalValue::dropLLVMManglingEscape(GV->getName()));
// Emit FPO data, but only on 32-bit x86. No other platforms use it.
if (Triple(MMI->getModule()->getTargetTriple()).getArch() == Triple::x86)
// Emit a symbol subsection, required by VS2012+ to find function boundaries.
OS.AddComment("Symbol subsection for " + Twine(FuncName));
MCSymbol *SymbolsEnd = beginCVSubsection(DebugSubsectionKind::Symbols);
SymbolKind ProcKind = GV->hasLocalLinkage() ? SymbolKind::S_LPROC32_ID
: SymbolKind::S_GPROC32_ID;
MCSymbol *ProcRecordEnd = beginSymbolRecord(ProcKind);
// These fields are filled in by tools like CVPACK which run after the fact.
// This is the important bit that tells the debugger where the function
// code is located and what's its size:
OS.AddComment("Code size");
OS.emitAbsoluteSymbolDiff(FI.End, Fn, 4);
OS.AddComment("Offset after prologue");
OS.AddComment("Offset before epilogue");
OS.AddComment("Function type index");
OS.AddComment("Function section relative address");
OS.EmitCOFFSecRel32(Fn, /*Offset=*/0);
OS.AddComment("Function section index");
// Emit the function display name as a null-terminated string.
OS.AddComment("Function name");
// Truncate the name so we won't overflow the record length field.
emitNullTerminatedSymbolName(OS, FuncName);
MCSymbol *FrameProcEnd = beginSymbolRecord(SymbolKind::S_FRAMEPROC);
// Subtract out the CSR size since MSVC excludes that and we include it.
OS.emitInt32(FI.FrameSize - FI.CSRSize);
OS.AddComment("Offset of padding");
OS.AddComment("Bytes of callee saved registers");
OS.AddComment("Exception handler offset");
OS.AddComment("Exception handler section");
OS.AddComment("Flags (defines frame register)");
emitLocalVariableList(FI, FI.Locals);
emitLexicalBlockList(FI.ChildBlocks, FI);
// Emit inlined call site information. Only emit functions inlined directly
// into the parent function. We'll emit the other sites recursively as part
// of their parent inline site.
for (const DILocation *InlinedAt : FI.ChildSites) {
auto I = FI.InlineSites.find(InlinedAt);
assert(I != FI.InlineSites.end() &&
"child site not in function inline site map");
emitInlinedCallSite(FI, InlinedAt, I->second);
for (auto Annot : FI.Annotations) {
MCSymbol *Label = Annot.first;
MDTuple *Strs = cast<MDTuple>(Annot.second);
MCSymbol *AnnotEnd = beginSymbolRecord(SymbolKind::S_ANNOTATION);
OS.EmitCOFFSecRel32(Label, /*Offset=*/0);
// FIXME: Make sure we don't overflow the max record size.
for (Metadata *MD : Strs->operands()) {
// MDStrings are null terminated, so we can do EmitBytes and get the
// nice .asciz directive.
StringRef Str = cast<MDString>(MD)->getString();
assert([Str.size()] == '\0' && "non-nullterminated MDString");
OS.emitBytes(StringRef(, Str.size() + 1));
for (auto HeapAllocSite : FI.HeapAllocSites) {
const MCSymbol *BeginLabel = std::get<0>(HeapAllocSite);
const MCSymbol *EndLabel = std::get<1>(HeapAllocSite);
const DIType *DITy = std::get<2>(HeapAllocSite);
MCSymbol *HeapAllocEnd = beginSymbolRecord(SymbolKind::S_HEAPALLOCSITE);
OS.AddComment("Call site offset");
OS.EmitCOFFSecRel32(BeginLabel, /*Offset=*/0);
OS.AddComment("Call site section index");
OS.AddComment("Call instruction length");
OS.emitAbsoluteSymbolDiff(EndLabel, BeginLabel, 2);
OS.AddComment("Type index");
if (SP != nullptr)
// We're done with this function.
// We have an assembler directive that takes care of the whole line table.
OS.emitCVLinetableDirective(FI.FuncId, Fn, FI.End);
CodeViewDebug::createDefRangeMem(uint16_t CVRegister, int Offset) {
LocalVarDefRange DR;
DR.InMemory = -1;
DR.DataOffset = Offset;
assert(DR.DataOffset == Offset && "truncation");
DR.IsSubfield = 0;
DR.StructOffset = 0;
DR.CVRegister = CVRegister;
return DR;
void CodeViewDebug::collectVariableInfoFromMFTable(
DenseSet<InlinedEntity> &Processed) {
const MachineFunction &MF = *Asm->MF;
const TargetSubtargetInfo &TSI = MF.getSubtarget();
const TargetFrameLowering *TFI = TSI.getFrameLowering();
const TargetRegisterInfo *TRI = TSI.getRegisterInfo();
for (const MachineFunction::VariableDbgInfo &VI : MF.getVariableDbgInfo()) {
if (!VI.Var)
assert(VI.Var->isValidLocationForIntrinsic(VI.Loc) &&
"Expected inlined-at fields to agree");
Processed.insert(InlinedEntity(VI.Var, VI.Loc->getInlinedAt()));
LexicalScope *Scope = LScopes.findLexicalScope(VI.Loc);
// If variable scope is not found then skip this variable.
if (!Scope)
// If the variable has an attached offset expression, extract it.
// FIXME: Try to handle DW_OP_deref as well.
int64_t ExprOffset = 0;
bool Deref = false;
if (VI.Expr) {
// If there is one DW_OP_deref element, use offset of 0 and keep going.
if (VI.Expr->getNumElements() == 1 &&
VI.Expr->getElement(0) == llvm::dwarf::DW_OP_deref)
Deref = true;
else if (!VI.Expr->extractIfOffset(ExprOffset))
// Get the frame register used and the offset.
Register FrameReg;
StackOffset FrameOffset = TFI->getFrameIndexReference(*Asm->MF, VI.Slot, FrameReg);
uint16_t CVReg = TRI->getCodeViewRegNum(FrameReg);
assert(!FrameOffset.getScalable() &&
"Frame offsets with a scalable component are not supported");
// Calculate the label ranges.
LocalVarDefRange DefRange =
createDefRangeMem(CVReg, FrameOffset.getFixed() + ExprOffset);
for (const InsnRange &Range : Scope->getRanges()) {
const MCSymbol *Begin = getLabelBeforeInsn(Range.first);
const MCSymbol *End = getLabelAfterInsn(Range.second);
End = End ? End : Asm->getFunctionEnd();
DefRange.Ranges.emplace_back(Begin, End);
LocalVariable Var;
Var.DIVar = VI.Var;
if (Deref)
Var.UseReferenceType = true;
recordLocalVariable(std::move(Var), Scope);
static bool canUseReferenceType(const DbgVariableLocation &Loc) {
return !Loc.LoadChain.empty() && Loc.LoadChain.back() == 0;
static bool needsReferenceType(const DbgVariableLocation &Loc) {
return Loc.LoadChain.size() == 2 && Loc.LoadChain.back() == 0;
void CodeViewDebug::calculateRanges(
LocalVariable &Var, const DbgValueHistoryMap::Entries &Entries) {
const TargetRegisterInfo *TRI = Asm->MF->getSubtarget().getRegisterInfo();
// Calculate the definition ranges.
for (auto I = Entries.begin(), E = Entries.end(); I != E; ++I) {
const auto &Entry = *I;
if (!Entry.isDbgValue())
const MachineInstr *DVInst = Entry.getInstr();
assert(DVInst->isDebugValue() && "Invalid History entry");
// FIXME: Find a way to represent constant variables, since they are
// relatively common.
Optional<DbgVariableLocation> Location =
if (!Location)
// CodeView can only express variables in register and variables in memory
// at a constant offset from a register. However, for variables passed
// indirectly by pointer, it is common for that pointer to be spilled to a
// stack location. For the special case of one offseted load followed by a
// zero offset load (a pointer spilled to the stack), we change the type of
// the local variable from a value type to a reference type. This tricks the
// debugger into doing the load for us.
if (Var.UseReferenceType) {
// We're using a reference type. Drop the last zero offset load.
if (canUseReferenceType(*Location))
} else if (needsReferenceType(*Location)) {
// This location can't be expressed without switching to a reference type.
// Start over using that.
Var.UseReferenceType = true;
calculateRanges(Var, Entries);
// We can only handle a register or an offseted load of a register.
if (Location->Register == 0 || Location->LoadChain.size() > 1)
LocalVarDefRange DR;
DR.CVRegister = TRI->getCodeViewRegNum(Location->Register);
DR.InMemory = !Location->LoadChain.empty();
DR.DataOffset =
!Location->LoadChain.empty() ? Location->LoadChain.back() : 0;
if (Location->FragmentInfo) {
DR.IsSubfield = true;
DR.StructOffset = Location->FragmentInfo->OffsetInBits / 8;
} else {
DR.IsSubfield = false;
DR.StructOffset = 0;
if (Var.DefRanges.empty() ||
Var.DefRanges.back().isDifferentLocation(DR)) {
// Compute the label range.
const MCSymbol *Begin = getLabelBeforeInsn(Entry.getInstr());
const MCSymbol *End;
if (Entry.getEndIndex() != DbgValueHistoryMap::NoEntry) {
auto &EndingEntry = Entries[Entry.getEndIndex()];
End = EndingEntry.isDbgValue()
? getLabelBeforeInsn(EndingEntry.getInstr())
: getLabelAfterInsn(EndingEntry.getInstr());
} else
End = Asm->getFunctionEnd();
// If the last range end is our begin, just extend the last range.
// Otherwise make a new range.
SmallVectorImpl<std::pair<const MCSymbol *, const MCSymbol *>> &R =
if (!R.empty() && R.back().second == Begin)
R.back().second = End;
R.emplace_back(Begin, End);
// FIXME: Do more range combining.
void CodeViewDebug::collectVariableInfo(const DISubprogram *SP) {
DenseSet<InlinedEntity> Processed;
// Grab the variable info that was squirreled away in the MMI side-table.
for (const auto &I : DbgValues) {
InlinedEntity IV = I.first;
if (Processed.count(IV))
const DILocalVariable *DIVar = cast<DILocalVariable>(IV.first);
const DILocation *InlinedAt = IV.second;
// Instruction ranges, specifying where IV is accessible.
const auto &Entries = I.second;
LexicalScope *Scope = nullptr;
if (InlinedAt)
Scope = LScopes.findInlinedScope(DIVar->getScope(), InlinedAt);
Scope = LScopes.findLexicalScope(DIVar->getScope());
// If variable scope is not found then skip this variable.
if (!Scope)
LocalVariable Var;
Var.DIVar = DIVar;
calculateRanges(Var, Entries);
recordLocalVariable(std::move(Var), Scope);
void CodeViewDebug::beginFunctionImpl(const MachineFunction *MF) {
const TargetSubtargetInfo &TSI = MF->getSubtarget();
const TargetRegisterInfo *TRI = TSI.getRegisterInfo();
const MachineFrameInfo &MFI = MF->getFrameInfo();
const Function &GV = MF->getFunction();
auto Insertion = FnDebugInfo.insert({&GV, std::make_unique<FunctionInfo>()});
assert(Insertion.second && "function already has info");
CurFn = Insertion.first->second.get();
CurFn->FuncId = NextFuncId++;
CurFn->Begin = Asm->getFunctionBegin();
// The S_FRAMEPROC record reports the stack size, and how many bytes of
// callee-saved registers were used. For targets that don't use a PUSH
// instruction (AArch64), this will be zero.
CurFn->CSRSize = MFI.getCVBytesOfCalleeSavedRegisters();
CurFn->FrameSize = MFI.getStackSize();
CurFn->OffsetAdjustment = MFI.getOffsetAdjustment();
CurFn->HasStackRealignment = TRI->needsStackRealignment(*MF);
// For this function S_FRAMEPROC record, figure out which codeview register
// will be the frame pointer.
CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::None; // None.
CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::None; // None.
if (CurFn->FrameSize > 0) {
if (!TSI.getFrameLowering()->hasFP(*MF)) {
CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr;
CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::StackPtr;
} else {
// If there is an FP, parameters are always relative to it.
CurFn->EncodedParamFramePtrReg = EncodedFramePtrReg::FramePtr;
if (CurFn->HasStackRealignment) {
// If the stack needs realignment, locals are relative to SP or VFRAME.
CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::StackPtr;
} else {
// Otherwise, locals are relative to EBP, and we probably have VLAs or
// other stack adjustments.
CurFn->EncodedLocalFramePtrReg = EncodedFramePtrReg::FramePtr;
// Compute other frame procedure options.
FrameProcedureOptions FPO = FrameProcedureOptions::None;
if (MFI.hasVarSizedObjects())
FPO |= FrameProcedureOptions::HasAlloca;
if (MF->exposesReturnsTwice())
FPO |= FrameProcedureOptions::HasSetJmp;
// FIXME: Set HasLongJmp if we ever track that info.
if (MF->hasInlineAsm())
FPO |= FrameProcedureOptions::HasInlineAssembly;
if (GV.hasPersonalityFn()) {
if (isAsynchronousEHPersonality(
FPO |= FrameProcedureOptions::HasStructuredExceptionHandling;
FPO |= FrameProcedureOptions::HasExceptionHandling;
if (GV.hasFnAttribute(Attribute::InlineHint))
FPO |= FrameProcedureOptions::MarkedInline;
if (GV.hasFnAttribute(Attribute::Naked))
FPO |= FrameProcedureOptions::Naked;
if (MFI.hasStackProtectorIndex())
FPO |= FrameProcedureOptions::SecurityChecks;
FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedLocalFramePtrReg) << 14U);
FPO |= FrameProcedureOptions(uint32_t(CurFn->EncodedParamFramePtrReg) << 16U);
if (Asm->TM.getOptLevel() != CodeGenOpt::None &&
!GV.hasOptSize() && !GV.hasOptNone())
FPO |= FrameProcedureOptions::OptimizedForSpeed;
// FIXME: Set GuardCfg when it is implemented.
CurFn->FrameProcOpts = FPO;
// Find the end of the function prolog. First known non-DBG_VALUE and
// non-frame setup location marks the beginning of the function body.
// FIXME: is there a simpler a way to do this? Can we just search
// for the first instruction of the function, not the last of the prolog?
DebugLoc PrologEndLoc;
bool EmptyPrologue = true;
for (const auto &MBB : *MF) {
for (const auto &MI : MBB) {
if (!MI.isMetaInstruction() && !MI.getFlag(MachineInstr::FrameSetup) &&
MI.getDebugLoc()) {
PrologEndLoc = MI.getDebugLoc();
} else if (!MI.isMetaInstruction()) {
EmptyPrologue = false;
// Record beginning of function if we have a non-empty prologue.
if (PrologEndLoc && !EmptyPrologue) {
DebugLoc FnStartDL = PrologEndLoc.getFnDebugLoc();
maybeRecordLocation(FnStartDL, MF);
// Find heap alloc sites and emit labels around them.
for (const auto &MBB : *MF) {
for (const auto &MI : MBB) {
if (MI.getHeapAllocMarker()) {
static bool shouldEmitUdt(const DIType *T) {
if (!T)
return false;
// MSVC does not emit UDTs for typedefs that are scoped to classes.
if (T->getTag() == dwarf::DW_TAG_typedef) {
if (DIScope *Scope = T->getScope()) {
switch (Scope->getTag()) {
case dwarf::DW_TAG_structure_type:
case dwarf::DW_TAG_class_type:
case dwarf::DW_TAG_union_type:
return false;
while (true) {
if (!T || T->isForwardDecl())
return false;
const DIDerivedType *DT = dyn_cast<DIDerivedType>(T);
if (!DT)
return true;
T = DT->getBaseType();
return true;
void CodeViewDebug::addToUDTs(const DIType *Ty) {
// Don't record empty UDTs.
if (Ty->getName().empty())
if (!shouldEmitUdt(Ty))
SmallVector<StringRef, 5> ParentScopeNames;
const DISubprogram *ClosestSubprogram =
collectParentScopeNames(Ty->getScope(), ParentScopeNames);
std::string FullyQualifiedName =
formatNestedName(ParentScopeNames, getPrettyScopeName(Ty));
if (ClosestSubprogram == nullptr) {
GlobalUDTs.emplace_back(std::move(FullyQualifiedName), Ty);
} else if (ClosestSubprogram == CurrentSubprogram) {
LocalUDTs.emplace_back(std::move(FullyQualifiedName), Ty);
// TODO: What if the ClosestSubprogram is neither null or the current
// subprogram? Currently, the UDT just gets dropped on the floor.
// The current behavior is not desirable. To get maximal fidelity, we would
// need to perform all type translation before beginning emission of .debug$S
// and then make LocalUDTs a member of FunctionInfo
TypeIndex CodeViewDebug::lowerType(const DIType *Ty, const DIType *ClassTy) {
// Generic dispatch for lowering an unknown type.
switch (Ty->getTag()) {
case dwarf::DW_TAG_array_type:
return lowerTypeArray(cast<DICompositeType>(Ty));
case dwarf::DW_TAG_typedef:
return lowerTypeAlias(cast<DIDerivedType>(Ty));
case dwarf::DW_TAG_base_type:
return lowerTypeBasic(cast<DIBasicType>(Ty));
case dwarf::DW_TAG_pointer_type:
if (cast<DIDerivedType>(Ty)->getName() == "__vtbl_ptr_type")
return lowerTypeVFTableShape(cast<DIDerivedType>(Ty));
case dwarf::DW_TAG_reference_type:
case dwarf::DW_TAG_rvalue_reference_type:
return lowerTypePointer(cast<DIDerivedType>(Ty));
case dwarf::DW_TAG_ptr_to_member_type:
return lowerTypeMemberPointer(cast<DIDerivedType>(Ty));
case dwarf::DW_TAG_restrict_type:
case dwarf::DW_TAG_const_type:
case dwarf::DW_TAG_volatile_type:
// TODO: add support for DW_TAG_atomic_type here
return lowerTypeModifier(cast<DIDerivedType>(Ty));
case dwarf::DW_TAG_subroutine_type:
if (ClassTy) {
// The member function type of a member function pointer has no
// ThisAdjustment.
return lowerTypeMemberFunction(cast<DISubroutineType>(Ty), ClassTy,
return lowerTypeFunction(cast<DISubroutineType>(Ty));
case dwarf::DW_TAG_enumeration_type:
return lowerTypeEnum(cast<DICompositeType>(Ty));
case dwarf::DW_TAG_class_type:
case dwarf::DW_TAG_structure_type:
return lowerTypeClass(cast<DICompositeType>(Ty));
case dwarf::DW_TAG_union_type:
return lowerTypeUnion(cast<DICompositeType>(Ty));
case dwarf::DW_TAG_unspecified_type:
if (Ty->getName() == "decltype(nullptr)")
return TypeIndex::NullptrT();
return TypeIndex::None();
// Use the null type index.
return TypeIndex();
TypeIndex CodeViewDebug::lowerTypeAlias(const DIDerivedType *Ty) {
TypeIndex UnderlyingTypeIndex = getTypeIndex(Ty->getBaseType());
StringRef TypeName = Ty->getName();
if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::Int32Long) &&
TypeName == "HRESULT")
return TypeIndex(SimpleTypeKind::HResult);
if (UnderlyingTypeIndex == TypeIndex(SimpleTypeKind::UInt16Short) &&
TypeName == "wchar_t")
return TypeIndex(SimpleTypeKind::WideCharacter);
return UnderlyingTypeIndex;
TypeIndex CodeViewDebug::lowerTypeArray(const DICompositeType *Ty) {
const DIType *ElementType = Ty->getBaseType();
TypeIndex ElementTypeIndex = getTypeIndex(ElementType);
// IndexType is size_t, which depends on the bitness of the target.
TypeIndex IndexType = getPointerSizeInBytes() == 8
? TypeIndex(SimpleTypeKind::UInt64Quad)
: TypeIndex(SimpleTypeKind::UInt32Long);
uint64_t ElementSize = getBaseTypeSize(ElementType) / 8;
// Add subranges to array type.
DINodeArray Elements = Ty->getElements();
for (int i = Elements.size() - 1; i >= 0; --i) {
const DINode *Element = Elements[i];
assert(Element->getTag() == dwarf::DW_TAG_subrange_type);
const DISubrange *Subrange = cast<DISubrange>(Element);
int64_t Count = -1;
// Calculate the count if either LowerBound is absent or is zero and
// either of Count or UpperBound are constant.
auto *LI = Subrange->getLowerBound().dyn_cast<ConstantInt *>();
if (!Subrange->getRawLowerBound() || (LI && (LI->getSExtValue() == 0))) {
if (auto *CI = Subrange->getCount().dyn_cast<ConstantInt*>())
Count = CI->getSExtValue();
else if (auto *UI = Subrange->getUpperBound().dyn_cast<ConstantInt*>())
Count = UI->getSExtValue() + 1; // LowerBound is zero
// Forward declarations of arrays without a size and VLAs use a count of -1.
// Emit a count of zero in these cases to match what MSVC does for arrays
// without a size. MSVC doesn't support VLAs, so it's not clear what we
// should do for them even if we could distinguish them.
if (Count == -1)
Count = 0;
// Update the element size and element type index for subsequent subranges.
ElementSize *= Count;
// If this is the outermost array, use the size from the array. It will be
// more accurate if we had a VLA or an incomplete element type size.
uint64_t ArraySize =
(i == 0 && ElementSize == 0) ? Ty->getSizeInBits() / 8 : ElementSize;
StringRef Name = (i == 0) ? Ty->getName() : "";
ArrayRecord AR(ElementTypeIndex, IndexType, ArraySize, Name);
ElementTypeIndex = TypeTable.writeLeafType(AR);
return ElementTypeIndex;
TypeIndex CodeViewDebug::lowerTypeBasic(const DIBasicType *Ty) {
TypeIndex Index;
dwarf::TypeKind Kind;
uint32_t ByteSize;
Kind = static_cast<dwarf::TypeKind>(Ty->getEncoding());
ByteSize = Ty->getSizeInBits() / 8;
SimpleTypeKind STK = SimpleTypeKind::None;
switch (Kind) {
case dwarf::DW_ATE_address:
// FIXME: Translate
case dwarf::DW_ATE_boolean:
switch (ByteSize) {
case 1: STK = SimpleTypeKind::Boolean8; break;
case 2: STK = SimpleTypeKind::Boolean16; break;
case 4: STK = SimpleTypeKind::Boolean32; break;
case 8: STK = SimpleTypeKind::Boolean64; break;
case 16: STK = SimpleTypeKind::Boolean128; break;
case dwarf::DW_ATE_complex_float:
switch (ByteSize) {
case 2: STK = SimpleTypeKind::Complex16; break;
case 4: STK = SimpleTypeKind::Complex32; break;
case 8: STK = SimpleTypeKind::Complex64; break;
case 10: STK = SimpleTypeKind::Complex80; break;
case 16: STK = SimpleTypeKind::Complex128; break;
case dwarf::DW_ATE_float:
switch (ByteSize) {
case 2: STK = SimpleTypeKind::Float16; break;
case 4: STK = SimpleTypeKind::Float32; break;
case 6: STK = SimpleTypeKind::Float48; break;
case 8: STK = SimpleTypeKind::Float64; break;
case 10: STK = SimpleTypeKind::Float80; break;
case 16: STK = SimpleTypeKind::Float128; break;
case dwarf::DW_ATE_signed:
switch (ByteSize) {
case 1: STK = SimpleTypeKind::SignedCharacter; break;
case 2: STK = SimpleTypeKind::Int16Short; break;
case 4: STK = SimpleTypeKind::Int32; break;
case 8: STK = SimpleTypeKind::Int64Quad; break;
case 16: STK = SimpleTypeKind::Int128Oct; break;
case dwarf::DW_ATE_unsigned:
switch (ByteSize) {
case 1: STK = SimpleTypeKind::UnsignedCharacter; break;
case 2: STK = SimpleTypeKind::UInt16Short; break;
case 4: STK = SimpleTypeKind::UInt32; break;
case 8: STK = SimpleTypeKind::UInt64Quad; break;
case 16: STK = SimpleTypeKind::UInt128Oct; break;
case dwarf::DW_ATE_UTF:
switch (ByteSize) {
case 2: STK = SimpleTypeKind::Character16; break;
case 4: STK = SimpleTypeKind::Character32; break;
case dwarf::DW_ATE_signed_char:
if (ByteSize == 1)
STK = SimpleTypeKind::SignedCharacter;
case dwarf::DW_ATE_unsigned_char:
if (ByteSize == 1)
STK = SimpleTypeKind::UnsignedCharacter;
// Apply some fixups based on the source-level type name.
if (STK == SimpleTypeKind::Int32 && Ty->getName() == "long int")
STK = SimpleTypeKind::Int32Long;
if (STK == SimpleTypeKind::UInt32 && Ty->getName() == "long unsigned int")
STK = SimpleTypeKind::UInt32Long;
if (STK == SimpleTypeKind::UInt16Short &&
(Ty->getName() == "wchar_t" || Ty->getName() == "__wchar_t"))
STK = SimpleTypeKind::WideCharacter;
if ((STK == SimpleTypeKind::SignedCharacter ||
STK == SimpleTypeKind::UnsignedCharacter) &&
Ty->getName() == "char")
STK = SimpleTypeKind::NarrowCharacter;
return TypeIndex(STK);
TypeIndex CodeViewDebug::lowerTypePointer(const DIDerivedType *Ty,
PointerOptions PO) {
TypeIndex PointeeTI = getTypeIndex(Ty->getBaseType());
// Pointers to simple types without any options can use SimpleTypeMode, rather
// than having a dedicated pointer type record.
if (PointeeTI.isSimple() && PO == PointerOptions::None &&
PointeeTI.getSimpleMode() == SimpleTypeMode::Direct &&
Ty->getTag() == dwarf::DW_TAG_pointer_type) {
SimpleTypeMode Mode = Ty->getSizeInBits() == 64
? SimpleTypeMode::NearPointer64
: SimpleTypeMode::NearPointer32;
return TypeIndex(PointeeTI.getSimpleKind(), Mode);
PointerKind PK =
Ty->getSizeInBits() == 64 ? PointerKind::Near64 : PointerKind::Near32;
PointerMode PM = PointerMode::Pointer;
switch (Ty->getTag()) {
default: llvm_unreachable("not a pointer tag type");
case dwarf::DW_TAG_pointer_type:
PM = PointerMode::Pointer;
case dwarf::DW_TAG_reference_type:
PM = PointerMode::LValueReference;
case dwarf::DW_TAG_rvalue_reference_type:
PM = PointerMode::RValueReference;
if (Ty->isObjectPointer())
PO |= PointerOptions::Const;
PointerRecord PR(PointeeTI, PK, PM, PO, Ty->getSizeInBits() / 8);
return TypeTable.writeLeafType(PR);
static PointerToMemberRepresentation
translatePtrToMemberRep(unsigned SizeInBytes, bool IsPMF, unsigned Flags) {
// SizeInBytes being zero generally implies that the member pointer type was
// incomplete, which can happen if it is part of a function prototype. In this
// case, use the unknown model instead of the general model.
if (IsPMF) {
switch (Flags & DINode::FlagPtrToMemberRep) {
case 0:
return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown
: PointerToMemberRepresentation::GeneralFunction;
case DINode::FlagSingleInheritance:
return PointerToMemberRepresentation::SingleInheritanceFunction;
case DINode::FlagMultipleInheritance:
return PointerToMemberRepresentation::MultipleInheritanceFunction;
case DINode::FlagVirtualInheritance:
return PointerToMemberRepresentation::VirtualInheritanceFunction;
} else {
switch (Flags & DINode::FlagPtrToMemberRep) {
case 0:
return SizeInBytes == 0 ? PointerToMemberRepresentation::Unknown
: PointerToMemberRepresentation::GeneralData;
case DINode::FlagSingleInheritance:
return PointerToMemberRepresentation::SingleInheritanceData;
case DINode::FlagMultipleInheritance:
return PointerToMemberRepresentation::MultipleInheritanceData;
case DINode::FlagVirtualInheritance:
return PointerToMemberRepresentation::VirtualInheritanceData;
llvm_unreachable("invalid ptr to member representation");
TypeIndex CodeViewDebug::lowerTypeMemberPointer(const DIDerivedType *Ty,
PointerOptions PO) {
assert(Ty->getTag() == dwarf::DW_TAG_ptr_to_member_type);
bool IsPMF = isa<DISubroutineType>(Ty->getBaseType());
TypeIndex ClassTI = getTypeIndex(Ty->getClassType());
TypeIndex PointeeTI =
getTypeIndex(Ty->getBaseType(), IsPMF ? Ty->getClassType() : nullptr);
PointerKind PK = getPointerSizeInBytes() == 8 ? PointerKind::Near64
: PointerKind::Near32;
PointerMode PM = IsPMF ? PointerMode::PointerToMemberFunction
: PointerMode::PointerToDataMember;
assert(Ty->getSizeInBits() / 8 <= 0xff && "pointer size too big");
uint8_t SizeInBytes = Ty->getSizeInBits() / 8;
MemberPointerInfo MPI(
ClassTI, translatePtrToMemberRep(SizeInBytes, IsPMF, Ty->getFlags()));
PointerRecord PR(PointeeTI, PK, PM, PO, SizeInBytes, MPI);
return TypeTable.writeLeafType(PR);
/// Given a DWARF calling convention, get the CodeView equivalent. If we don't
/// have a translation, use the NearC convention.
static CallingConvention dwarfCCToCodeView(unsigned DwarfCC) {
switch (DwarfCC) {
case dwarf::DW_CC_normal: return CallingConvention::NearC;
case dwarf::DW_CC_BORLAND_msfastcall: return CallingConvention::NearFast;
case dwarf::DW_CC_BORLAND_thiscall: return CallingConvention::ThisCall;
case dwarf::DW_CC_BORLAND_stdcall: return CallingConvention::NearStdCall;
case dwarf::DW_CC_BORLAND_pascal: return CallingConvention::NearPascal;
case dwarf::DW_CC_LLVM_vectorcall: return CallingConvention::NearVector;
return CallingConvention::NearC;
TypeIndex CodeViewDebug::lowerTypeModifier(const DIDerivedType *Ty) {
ModifierOptions Mods = ModifierOptions::None;
PointerOptions PO = PointerOptions::None;
bool IsModifier = true;
const DIType *BaseTy = Ty;
while (IsModifier && BaseTy) {
// FIXME: Need to add DWARF tags for __unaligned and _Atomic
switch (BaseTy->getTag()) {
case dwarf::DW_TAG_const_type:
Mods |= ModifierOptions::Const;
PO |= PointerOptions::Const;
case dwarf::DW_TAG_volatile_type:
Mods |= ModifierOptions::Volatile;
PO |= PointerOptions::Volatile;
case dwarf::DW_TAG_restrict_type:
// Only pointer types be marked with __restrict. There is no known flag
// for __restrict in LF_MODIFIER records.
PO |= PointerOptions::Restrict;
IsModifier = false;
if (IsModifier)
BaseTy = cast<DIDerivedType>(BaseTy)->getBaseType();
// Check if the inner type will use an LF_POINTER record. If so, the
// qualifiers will go in the LF_POINTER record. This comes up for types like
// 'int *const' and 'int *__restrict', not the more common cases like 'const
// char *'.
if (BaseTy) {
switch (BaseTy->getTag()) {
case dwarf::DW_TAG_pointer_type:
case dwarf::DW_TAG_reference_type:
case dwarf::DW_TAG_rvalue_reference_type:
return lowerTypePointer(cast<DIDerivedType>(BaseTy), PO);
case dwarf::DW_TAG_ptr_to_member_type:
return lowerTypeMemberPointer(cast<DIDerivedType>(BaseTy), PO);
TypeIndex ModifiedTI = getTypeIndex(BaseTy);
// Return the base type index if there aren't any modifiers. For example, the
// metadata could contain restrict wrappers around non-pointer types.
if (Mods == ModifierOptions::None)
return ModifiedTI;
ModifierRecord MR(ModifiedTI, Mods);
return TypeTable.writeLeafType(MR);
TypeIndex CodeViewDebug::lowerTypeFunction(const DISubroutineType *Ty) {
SmallVector<TypeIndex, 8> ReturnAndArgTypeIndices;
for (const DIType *ArgType : Ty->getTypeArray())
// MSVC uses type none for variadic argument.
if (ReturnAndArgTypeIndices.size() > 1 &&
ReturnAndArgTypeIndices.back() == TypeIndex::Void()) {
ReturnAndArgTypeIndices.back() = TypeIndex::None();
TypeIndex ReturnTypeIndex = TypeIndex::Void();
ArrayRef<TypeIndex> ArgTypeIndices = None;
if (!ReturnAndArgTypeIndices.empty()) {
auto ReturnAndArgTypesRef = makeArrayRef(ReturnAndArgTypeIndices);
ReturnTypeIndex = ReturnAndArgTypesRef.front();
ArgTypeIndices = ReturnAndArgTypesRef.drop_front();
ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices);
TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec);
CallingConvention CC = dwarfCCToCodeView(Ty->getCC());
FunctionOptions FO = getFunctionOptions(Ty);
ProcedureRecord Procedure(ReturnTypeIndex, CC, FO, ArgTypeIndices.size(),
return TypeTable.writeLeafType(Procedure);
TypeIndex CodeViewDebug::lowerTypeMemberFunction(const DISubroutineType *Ty,
const DIType *ClassTy,
int ThisAdjustment,
bool IsStaticMethod,
FunctionOptions FO) {
// Lower the containing class type.
TypeIndex ClassType = getTypeIndex(ClassTy);
DITypeRefArray ReturnAndArgs = Ty->getTypeArray();
unsigned Index = 0;
SmallVector<TypeIndex, 8> ArgTypeIndices;
TypeIndex ReturnTypeIndex = TypeIndex::Void();
if (ReturnAndArgs.size() > Index) {
ReturnTypeIndex = getTypeIndex(ReturnAndArgs[Index++]);
// If the first argument is a pointer type and this isn't a static method,
// treat it as the special 'this' parameter, which is encoded separately from
// the arguments.
TypeIndex ThisTypeIndex;
if (!IsStaticMethod && ReturnAndArgs.size() > Index) {
if (const DIDerivedType *PtrTy =
dyn_cast_or_null<DIDerivedType>(ReturnAndArgs[Index])) {
if (PtrTy->getTag() == dwarf::DW_TAG_pointer_type) {
ThisTypeIndex = getTypeIndexForThisPtr(PtrTy, Ty);
while (Index < ReturnAndArgs.size())
// MSVC uses type none for variadic argument.
if (!ArgTypeIndices.empty() && ArgTypeIndices.back() == TypeIndex::Void())
ArgTypeIndices.back() = TypeIndex::None();
ArgListRecord ArgListRec(TypeRecordKind::ArgList, ArgTypeIndices);
TypeIndex ArgListIndex = TypeTable.writeLeafType(ArgListRec);
CallingConvention CC = dwarfCCToCodeView(Ty->getCC());
MemberFunctionRecord MFR(ReturnTypeIndex, ClassType, ThisTypeIndex, CC, FO,
ArgTypeIndices.size(), ArgListIndex, ThisAdjustment);
return TypeTable.writeLeafType(MFR);
TypeIndex CodeViewDebug::lowerTypeVFTableShape(const DIDerivedType *Ty) {
unsigned VSlotCount =
Ty->getSizeInBits() / (8 * Asm->MAI->getCodePointerSize());
SmallVector<VFTableSlotKind, 4> Slots(VSlotCount, VFTableSlotKind::Near);
VFTableShapeRecord VFTSR(Slots);
return TypeTable.writeLeafType(VFTSR);
static MemberAccess translateAccessFlags(unsigned RecordTag, unsigned Flags) {
switch (Flags & DINode::FlagAccessibility) {
case DINode::FlagPrivate: return MemberAccess::Private;
case DINode::FlagPublic: return MemberAccess::Public;
case DINode::FlagProtected: return MemberAccess::Protected;
case 0:
// If there was no explicit access control, provide the default for the tag.
return RecordTag == dwarf::DW_TAG_class_type ? MemberAccess::Private
: MemberAccess::Public;
llvm_unreachable("access flags are exclusive");
static MethodOptions translateMethodOptionFlags(const DISubprogram *SP) {
if (SP->isArtificial())
return MethodOptions::CompilerGenerated;
// FIXME: Handle other MethodOptions.
return MethodOptions::None;
static MethodKind translateMethodKindFlags(const DISubprogram *SP,
bool Introduced) {
if (SP->getFlags() & DINode::FlagStaticMember)
return MethodKind::Static;
switch (SP->getVirtuality()) {
case dwarf::DW_VIRTUALITY_none:
case dwarf::DW_VIRTUALITY_virtual:
return Introduced ? MethodKind::IntroducingVirtual : MethodKind::Virtual;
case dwarf::DW_VIRTUALITY_pure_virtual:
return Introduced ? MethodKind::PureIntroducingVirtual
: MethodKind::PureVirtual;
llvm_unreachable("unhandled virtuality case");
return MethodKind::Vanilla;
static TypeRecordKind getRecordKind(const DICompositeType *Ty) {
switch (Ty->getTag()) {
case dwarf::DW_TAG_class_type: return TypeRecordKind::Class;
case dwarf::DW_TAG_structure_type: return TypeRecordKind::Struct;
llvm_unreachable("unexpected tag");
/// Return ClassOptions that should be present on both the forward declaration
/// and the defintion of a tag type.
static ClassOptions getCommonClassOptions(const DICompositeType *Ty) {
ClassOptions CO = ClassOptions::None;
// MSVC always sets this flag, even for local types. Clang doesn't always
// appear to give every type a linkage name, which may be problematic for us.
// FIXME: Investigate the consequences of not following them here.
if (!Ty->getIdentifier().empty())
CO |= ClassOptions::HasUniqueName;
// Put the Nested flag on a type if it appears immediately inside a tag type.
// Do not walk the scope chain. Do not attempt to compute ContainsNestedClass
// here. That flag is only set on definitions, and not forward declarations.
const DIScope *ImmediateScope = Ty->getScope();
if (ImmediateScope && isa<DICompositeType>(ImmediateScope))
CO |= ClassOptions::Nested;
// Put the Scoped flag on function-local types. MSVC puts this flag for enum
// type only when it has an immediate function scope. Clang never puts enums
// inside DILexicalBlock scopes. Enum types, as generated by clang, are
// always in function, class, or file scopes.
if (Ty->getTag() == dwarf::DW_TAG_enumeration_type) {
if (ImmediateScope && isa<DISubprogram>(ImmediateScope))
CO |= ClassOptions::Scoped;
} else {
for (const DIScope *Scope = ImmediateScope; Scope != nullptr;
Scope = Scope->getScope()) {
if (isa<DISubprogram>(Scope)) {
CO |= ClassOptions::Scoped;
return CO;
void CodeViewDebug::addUDTSrcLine(const DIType *Ty, TypeIndex TI) {
switch (Ty->getTag()) {
case dwarf::DW_TAG_class_type:
case dwarf::DW_TAG_structure_type:
case dwarf::DW_TAG_union_type:
case dwarf::DW_TAG_enumeration_type:
if (const auto *File = Ty->getFile()) {
StringIdRecord SIDR(TypeIndex(0x0), getFullFilepath(File));
TypeIndex SIDI = TypeTable.writeLeafType(SIDR);
UdtSourceLineRecord USLR(TI, SIDI, Ty->getLine());
TypeIndex CodeViewDebug::lowerTypeEnum(const DICompositeType *Ty) {
ClassOptions CO = getCommonClassOptions(Ty);
TypeIndex FTI;
unsigned EnumeratorCount = 0;
if (Ty->isForwardDecl()) {
CO |= ClassOptions::ForwardReference;
} else {
ContinuationRecordBuilder ContinuationBuilder;
for (const DINode *Element : Ty->getElements()) {
// We assume that the frontend provides all members in source declaration
// order, which is what MSVC does.
if (auto *Enumerator = dyn_cast_or_null<DIEnumerator>(Element)) {
EnumeratorRecord ER(MemberAccess::Public,
APSInt(Enumerator->getValue(), true),
FTI = TypeTable.insertRecord(ContinuationBuilder);
std::string FullName = getFullyQualifiedName(Ty);
EnumRecord ER(EnumeratorCount, CO, FTI, FullName, Ty->getIdentifier(),
TypeIndex EnumTI = TypeTable.writeLeafType(ER);
addUDTSrcLine(Ty, EnumTI);
return EnumTI;
// ClassInfo
struct llvm::ClassInfo {
struct MemberInfo {
const DIDerivedType *MemberTypeNode;
uint64_t BaseOffset;
// [MemberInfo]
using MemberList = std::vector<MemberInfo>;
using MethodsList = TinyPtrVector<const DISubprogram *>;
// MethodName -> MethodsList
using MethodsMap = MapVector<MDString *, MethodsList>;
/// Base classes.
std::vector<const DIDerivedType *> Inheritance;
/// Direct members.
MemberList Members;
// Direct overloaded methods gathered by name.
MethodsMap Methods;
TypeIndex VShapeTI;
std::vector<const DIType *> NestedTypes;
void CodeViewDebug::clear() {
assert(CurFn == nullptr);
void CodeViewDebug::collectMemberInfo(ClassInfo &Info,
const DIDerivedType *DDTy) {
if (!DDTy->getName().empty()) {
Info.Members.push_back({DDTy, 0});
// Collect static const data members with values.
if ((DDTy->getFlags() & DINode::FlagStaticMember) ==
DINode::FlagStaticMember) {
if (DDTy->getConstant() && (isa<ConstantInt>(DDTy->getConstant()) ||
// An unnamed member may represent a nested struct or union. Attempt to
// interpret the unnamed member as a DICompositeType possibly wrapped in
// qualifier types. Add all the indirect fields to the current record if that
// succeeds, and drop the member if that fails.
assert((DDTy->getOffsetInBits() % 8) == 0 && "Unnamed bitfield member!");
uint64_t Offset = DDTy->getOffsetInBits();
const DIType *Ty = DDTy->getBaseType();
bool FullyResolved = false;
while (!FullyResolved) {
switch (Ty->getTag()) {
case dwarf::DW_TAG_const_type:
case dwarf::DW_TAG_volatile_type:
// FIXME: we should apply the qualifier types to the indirect fields
// rather than dropping them.
Ty = cast<DIDerivedType>(Ty)->getBaseType();
FullyResolved = true;
const DICompositeType *DCTy = dyn_cast<DICompositeType>(Ty);
if (!DCTy)
ClassInfo NestedInfo = collectClassInfo(DCTy);
for (const ClassInfo::MemberInfo &IndirectField : NestedInfo.Members)
{IndirectField.MemberTypeNode, IndirectField.BaseOffset + Offset});
ClassInfo CodeViewDebug::collectClassInfo(const DICompositeType *Ty) {
ClassInfo Info;
// Add elements to structure type.
DINodeArray Elements = Ty->getElements();
for (auto *Element : Elements) {
// We assume that the frontend provides all members in source declaration
// order, which is what MSVC does.
if (!Element)
if (auto *SP = dyn_cast<DISubprogram>(Element)) {
} else if (auto *DDTy = dyn_cast<DIDerivedType>(Element)) {
if (DDTy->getTag() == dwarf::DW_TAG_member) {
collectMemberInfo(Info, DDTy);
} else if (DDTy->getTag() == dwarf::DW_TAG_inheritance) {
} else if (DDTy->getTag() == dwarf::DW_TAG_pointer_type &&
DDTy->getName() == "__vtbl_ptr_type") {
Info.VShapeTI = getTypeIndex(DDTy);
} else if (DDTy->getTag() == dwarf::DW_TAG_typedef) {
} else if (DDTy->getTag() == dwarf::DW_TAG_friend) {
// Ignore friend members. It appears that MSVC emitted info about
// friends in the past, but modern versions do not.
} else if (auto *Composite = dyn_cast<DICompositeType>(Element)) {
// Skip other unrecognized kinds of elements.
return Info;
static bool shouldAlwaysEmitCompleteClassType(const DICompositeType *Ty) {
// This routine is used by lowerTypeClass and lowerTypeUnion to determine
// if a complete type should be emitted instead of a forward reference.
return Ty->getName().empty() && Ty->getIdentifier().empty() &&
TypeIndex CodeViewDebug::lowerTypeClass(const DICompositeType *Ty) {
// Emit the complete type for unnamed structs. C++ classes with methods
// which have a circular reference back to the class type are expected to
// be named by the front-end and should not be "unnamed". C unnamed
// structs should not have circular references.
if (shouldAlwaysEmitCompleteClassType(Ty)) {
// If this unnamed complete type is already in the process of being defined
// then the description of the type is malformed and cannot be emitted
// into CodeView correctly so report a fatal error.
auto I = CompleteTypeIndices.find(Ty);
if (I != CompleteTypeIndices.end() && I->second == TypeIndex())
report_fatal_error("cannot debug circular reference to unnamed type");
return getCompleteTypeIndex(Ty);
// First, construct the forward decl. Don't look into Ty to compute the
// forward decl options, since it might not be available in all TUs.
TypeRecordKind Kind = getRecordKind(Ty);
ClassOptions CO =
ClassOptions::ForwardReference | getCommonClassOptions(Ty);
std::string FullName = getFullyQualifiedName(Ty);
ClassRecord CR(Kind, 0, CO, TypeIndex(), TypeIndex(), TypeIndex(), 0,
FullName, Ty->getIdentifier());
TypeIndex FwdDeclTI = TypeTable.writeLeafType(CR);
if (!Ty->isForwardDecl())
return FwdDeclTI;
TypeIndex CodeViewDebug::lowerCompleteTypeClass(const DICompositeType *Ty) {
// Construct the field list and complete type record.
TypeRecordKind Kind = getRecordKind(Ty);
ClassOptions CO = getCommonClassOptions(Ty);
TypeIndex FieldTI;
TypeIndex VShapeTI;
unsigned FieldCount;
bool ContainsNestedClass;
std::tie(FieldTI, VShapeTI, FieldCount, ContainsNestedClass) =
if (ContainsNestedClass)
CO |= ClassOptions::ContainsNestedClass;
// MSVC appears to set this flag by searching any destructor or method with
// FunctionOptions::Constructor among the emitted members. Clang AST has all
// the members, however special member functions are not yet emitted into
// debug information. For now checking a class's non-triviality seems enough.
// FIXME: not true for a nested unnamed struct.
if (isNonTrivial(Ty))
CO |= ClassOptions::HasConstructorOrDestructor;
std::string FullName = getFullyQualifiedName(Ty);
uint64_t SizeInBytes = Ty->getSizeInBits() / 8;
ClassRecord CR(Kind, FieldCount, CO, FieldTI, TypeIndex(), VShapeTI,
SizeInBytes, FullName, Ty->getIdentifier());
TypeIndex ClassTI = TypeTable.writeLeafType(CR);
addUDTSrcLine(Ty, ClassTI);
return ClassTI;
TypeIndex CodeViewDebug::lowerTypeUnion(const DICompositeType *Ty) {
// Emit the complete type for unnamed unions.
if (shouldAlwaysEmitCompleteClassType(Ty))
return getCompleteTypeIndex(Ty);
ClassOptions CO =
ClassOptions::ForwardReference | getCommonClassOptions(Ty);
std::string FullName = getFullyQualifiedName(Ty);
UnionRecord UR(0, CO, TypeIndex(), 0, FullName, Ty->getIdentifier());
TypeIndex FwdDeclTI = TypeTable.writeLeafType(UR);
if (!Ty->isForwardDecl())
return FwdDeclTI;
TypeIndex CodeViewDebug::lowerCompleteTypeUnion(const DICompositeType *Ty) {
ClassOptions CO = ClassOptions::Sealed | getCommonClassOptions(Ty);
TypeIndex FieldTI;
unsigned FieldCount;
bool ContainsNestedClass;
std::tie(FieldTI, std::ignore, FieldCount, ContainsNestedClass) =
if (ContainsNestedClass)
CO |= ClassOptions::ContainsNestedClass;
uint64_t SizeInBytes = Ty->getSizeInBits() / 8;
std::string FullName = getFullyQualifiedName(Ty);
UnionRecord UR(FieldCount, CO, FieldTI, SizeInBytes, FullName,
TypeIndex UnionTI = TypeTable.writeLeafType(UR);
addUDTSrcLine(Ty, UnionTI);
return UnionTI;
std::tuple<TypeIndex, TypeIndex, unsigned, bool>
CodeViewDebug::lowerRecordFieldList(const DICompositeType *Ty) {
// Manually count members. MSVC appears to count everything that generates a
// field list record. Each individual overload in a method overload group
// contributes to this count, even though the overload group is a single field
// list record.
unsigned MemberCount = 0;
ClassInfo Info = collectClassInfo(Ty);
ContinuationRecordBuilder ContinuationBuilder;
// Create base classes.
for (const DIDerivedType *I : Info.Inheritance) {
if (I->getFlags() & DINode::FlagVirtual) {
// Virtual base.
unsigned VBPtrOffset = I->getVBPtrOffset();
// FIXME: Despite the accessor name, the offset is really in bytes.
unsigned VBTableIndex = I->getOffsetInBits() / 4;
auto RecordKind = (I->getFlags() & DINode::FlagIndirectVirtualBase) == DINode::FlagIndirectVirtualBase
? TypeRecordKind::IndirectVirtualBaseClass
: TypeRecordKind::VirtualBaseClass;
VirtualBaseClassRecord VBCR(
RecordKind, translateAccessFlags(Ty->getTag(), I->getFlags()),
getTypeIndex(I->getBaseType()), getVBPTypeIndex(), VBPtrOffset,
} else {
assert(I->getOffsetInBits() % 8 == 0 &&
"bases must be on byte boundaries");
BaseClassRecord BCR(translateAccessFlags(Ty->getTag(), I->getFlags()),
I->getOffsetInBits() / 8);
// Create members.
for (ClassInfo::MemberInfo &MemberInfo : Info.Members) {
const DIDerivedType *Member = MemberInfo.MemberTypeNode;
TypeIndex MemberBaseType = getTypeIndex(Member->getBaseType());
StringRef MemberName = Member->getName();
MemberAccess Access =
translateAccessFlags(Ty->getTag(), Member->getFlags());
if (Member->isStaticMember()) {
StaticDataMemberRecord SDMR(Access, MemberBaseType, MemberName);
// Virtual function pointer member.
if ((Member->getFlags() & DINode::FlagArtificial) &&
Member->getName().startswith("_vptr$")) {
VFPtrRecord VFPR(getTypeIndex(Member->getBaseType()));
// Data member.
uint64_t MemberOffsetInBits =
Member->getOffsetInBits() + MemberInfo.BaseOffset;
if (Member->isBitField()) {
uint64_t StartBitOffset = MemberOffsetInBits;
if (const auto *CI =
dyn_cast_or_null<ConstantInt>(Member->getStorageOffsetInBits())) {
MemberOffsetInBits = CI->getZExtValue() + MemberInfo.BaseOffset;
StartBitOffset -= MemberOffsetInBits;
BitFieldRecord BFR(MemberBaseType, Member->getSizeInBits(),
MemberBaseType = TypeTable.writeLeafType(BFR);
uint64_t MemberOffsetInBytes = MemberOffsetInBits / 8;
DataMemberRecord DMR(Access, MemberBaseType, MemberOffsetInBytes,
// Create methods
for (auto &MethodItr : Info.Methods) {
StringRef Name = MethodItr.first->getString();