1 //===-- LLVMSymbolize.cpp -------------------------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // Implementation for LLVM symbolization library.
12 //===----------------------------------------------------------------------===//
14 #include "LLVMSymbolize.h"
15 #include "llvm/ADT/STLExtras.h"
16 #include "llvm/Config/config.h"
17 #include "llvm/DebugInfo/DWARF/DWARFContext.h"
18 #include "llvm/DebugInfo/PDB/PDB.h"
19 #include "llvm/DebugInfo/PDB/PDBContext.h"
20 #include "llvm/Object/ELFObjectFile.h"
21 #include "llvm/Object/MachO.h"
22 #include "llvm/Support/Casting.h"
23 #include "llvm/Support/Compression.h"
24 #include "llvm/Support/DataExtractor.h"
25 #include "llvm/Support/Errc.h"
26 #include "llvm/Support/FileSystem.h"
27 #include "llvm/Support/MemoryBuffer.h"
28 #include "llvm/Support/Path.h"
40 static bool error(std::error_code ec) {
43 errs() << "LLVMSymbolizer: error reading file: " << ec.message() << ".\n";
47 static DILineInfoSpecifier
48 getDILineInfoSpecifier(const LLVMSymbolizer::Options &Opts) {
49 return DILineInfoSpecifier(
50 DILineInfoSpecifier::FileLineInfoKind::AbsoluteFilePath,
54 ModuleInfo::ModuleInfo(ObjectFile *Obj, DIContext *DICtx)
55 : Module(Obj), DebugInfoContext(DICtx) {
56 std::unique_ptr<DataExtractor> OpdExtractor;
57 uint64_t OpdAddress = 0;
58 // Find the .opd (function descriptor) section if any, for big-endian
60 if (Module->getArch() == Triple::ppc64) {
61 for (section_iterator Section : Module->sections()) {
63 if (!error(Section->getName(Name)) && Name == ".opd") {
65 if (!error(Section->getContents(Data))) {
66 OpdExtractor.reset(new DataExtractor(Data, Module->isLittleEndian(),
67 Module->getBytesInAddress()));
68 OpdAddress = Section->getAddress();
74 for (const SymbolRef &Symbol : Module->symbols()) {
75 addSymbol(Symbol, OpdExtractor.get(), OpdAddress);
77 bool NoSymbolTable = (Module->symbol_begin() == Module->symbol_end());
78 if (NoSymbolTable && Module->isELF()) {
79 // Fallback to dynamic symbol table, if regular symbol table is stripped.
80 std::pair<symbol_iterator, symbol_iterator> IDyn =
81 getELFDynamicSymbolIterators(Module);
82 for (symbol_iterator si = IDyn.first, se = IDyn.second; si != se; ++si) {
83 addSymbol(*si, OpdExtractor.get(), OpdAddress);
88 void ModuleInfo::addSymbol(const SymbolRef &Symbol, DataExtractor *OpdExtractor,
89 uint64_t OpdAddress) {
90 SymbolRef::Type SymbolType;
91 if (error(Symbol.getType(SymbolType)))
93 if (SymbolType != SymbolRef::ST_Function && SymbolType != SymbolRef::ST_Data)
95 uint64_t SymbolAddress;
96 if (error(Symbol.getAddress(SymbolAddress)) ||
97 SymbolAddress == UnknownAddressOrSize)
100 // For big-endian PowerPC64 ELF, symbols in the .opd section refer to
101 // function descriptors. The first word of the descriptor is a pointer to
102 // the function's code.
103 // For the purposes of symbolization, pretend the symbol's address is that
104 // of the function's code, not the descriptor.
105 uint64_t OpdOffset = SymbolAddress - OpdAddress;
106 uint32_t OpdOffset32 = OpdOffset;
107 if (OpdOffset == OpdOffset32 &&
108 OpdExtractor->isValidOffsetForAddress(OpdOffset32))
109 SymbolAddress = OpdExtractor->getAddress(&OpdOffset32);
112 // Getting symbol size is linear for Mach-O files, so assume that symbol
113 // occupies the memory range up to the following symbol.
114 if (isa<MachOObjectFile>(Module))
117 SymbolSize = Symbol.getSize();
118 if (SymbolSize == UnknownAddressOrSize)
121 StringRef SymbolName;
122 if (error(Symbol.getName(SymbolName)))
124 // Mach-O symbol table names have leading underscore, skip it.
125 if (Module->isMachO() && SymbolName.size() > 0 && SymbolName[0] == '_')
126 SymbolName = SymbolName.drop_front();
127 // FIXME: If a function has alias, there are two entries in symbol table
128 // with same address size. Make sure we choose the correct one.
129 auto &M = SymbolType == SymbolRef::ST_Function ? Functions : Objects;
130 SymbolDesc SD = { SymbolAddress, SymbolSize };
131 M.insert(std::make_pair(SD, SymbolName));
134 bool ModuleInfo::getNameFromSymbolTable(SymbolRef::Type Type, uint64_t Address,
135 std::string &Name, uint64_t &Addr,
136 uint64_t &Size) const {
137 const auto &SymbolMap = Type == SymbolRef::ST_Function ? Functions : Objects;
138 if (SymbolMap.empty())
140 SymbolDesc SD = { Address, Address };
141 auto SymbolIterator = SymbolMap.upper_bound(SD);
142 if (SymbolIterator == SymbolMap.begin())
145 if (SymbolIterator->first.Size != 0 &&
146 SymbolIterator->first.Addr + SymbolIterator->first.Size <= Address)
148 Name = SymbolIterator->second.str();
149 Addr = SymbolIterator->first.Addr;
150 Size = SymbolIterator->first.Size;
154 DILineInfo ModuleInfo::symbolizeCode(
155 uint64_t ModuleOffset, const LLVMSymbolizer::Options &Opts) const {
157 if (DebugInfoContext) {
158 LineInfo = DebugInfoContext->getLineInfoForAddress(
159 ModuleOffset, getDILineInfoSpecifier(Opts));
161 // Override function name from symbol table if necessary.
162 if (Opts.PrintFunctions != FunctionNameKind::None && Opts.UseSymbolTable) {
163 std::string FunctionName;
164 uint64_t Start, Size;
165 if (getNameFromSymbolTable(SymbolRef::ST_Function, ModuleOffset,
166 FunctionName, Start, Size)) {
167 LineInfo.FunctionName = FunctionName;
173 DIInliningInfo ModuleInfo::symbolizeInlinedCode(
174 uint64_t ModuleOffset, const LLVMSymbolizer::Options &Opts) const {
175 DIInliningInfo InlinedContext;
177 if (DebugInfoContext) {
178 InlinedContext = DebugInfoContext->getInliningInfoForAddress(
179 ModuleOffset, getDILineInfoSpecifier(Opts));
181 // Make sure there is at least one frame in context.
182 if (InlinedContext.getNumberOfFrames() == 0) {
183 InlinedContext.addFrame(DILineInfo());
185 // Override the function name in lower frame with name from symbol table.
186 if (Opts.PrintFunctions != FunctionNameKind::None && Opts.UseSymbolTable) {
187 DIInliningInfo PatchedInlinedContext;
188 for (uint32_t i = 0, n = InlinedContext.getNumberOfFrames(); i < n; i++) {
189 DILineInfo LineInfo = InlinedContext.getFrame(i);
191 std::string FunctionName;
192 uint64_t Start, Size;
193 if (getNameFromSymbolTable(SymbolRef::ST_Function, ModuleOffset,
194 FunctionName, Start, Size)) {
195 LineInfo.FunctionName = FunctionName;
198 PatchedInlinedContext.addFrame(LineInfo);
200 InlinedContext = PatchedInlinedContext;
202 return InlinedContext;
205 bool ModuleInfo::symbolizeData(uint64_t ModuleOffset, std::string &Name,
206 uint64_t &Start, uint64_t &Size) const {
207 return getNameFromSymbolTable(SymbolRef::ST_Data, ModuleOffset, Name, Start,
211 const char LLVMSymbolizer::kBadString[] = "??";
213 std::string LLVMSymbolizer::symbolizeCode(const std::string &ModuleName,
214 uint64_t ModuleOffset) {
215 ModuleInfo *Info = getOrCreateModuleInfo(ModuleName);
217 return printDILineInfo(DILineInfo());
218 if (Opts.PrintInlining) {
219 DIInliningInfo InlinedContext =
220 Info->symbolizeInlinedCode(ModuleOffset, Opts);
221 uint32_t FramesNum = InlinedContext.getNumberOfFrames();
222 assert(FramesNum > 0);
224 for (uint32_t i = 0; i < FramesNum; i++) {
225 DILineInfo LineInfo = InlinedContext.getFrame(i);
226 Result += printDILineInfo(LineInfo);
230 DILineInfo LineInfo = Info->symbolizeCode(ModuleOffset, Opts);
231 return printDILineInfo(LineInfo);
234 std::string LLVMSymbolizer::symbolizeData(const std::string &ModuleName,
235 uint64_t ModuleOffset) {
236 std::string Name = kBadString;
239 if (Opts.UseSymbolTable) {
240 if (ModuleInfo *Info = getOrCreateModuleInfo(ModuleName)) {
241 if (Info->symbolizeData(ModuleOffset, Name, Start, Size) && Opts.Demangle)
242 Name = DemangleName(Name);
245 std::stringstream ss;
246 ss << Name << "\n" << Start << " " << Size << "\n";
250 void LLVMSymbolizer::flush() {
251 DeleteContainerSeconds(Modules);
252 ObjectPairForPathArch.clear();
253 ObjectFileForArch.clear();
256 // For Path="/path/to/foo" and Basename="foo" assume that debug info is in
257 // /path/to/foo.dSYM/Contents/Resources/DWARF/foo.
258 // For Path="/path/to/bar.dSYM" and Basename="foo" assume that debug info is in
259 // /path/to/bar.dSYM/Contents/Resources/DWARF/foo.
261 std::string getDarwinDWARFResourceForPath(
262 const std::string &Path, const std::string &Basename) {
263 SmallString<16> ResourceName = StringRef(Path);
264 if (sys::path::extension(Path) != ".dSYM") {
265 ResourceName += ".dSYM";
267 sys::path::append(ResourceName, "Contents", "Resources", "DWARF");
268 sys::path::append(ResourceName, Basename);
269 return ResourceName.str();
272 static bool checkFileCRC(StringRef Path, uint32_t CRCHash) {
273 ErrorOr<std::unique_ptr<MemoryBuffer>> MB =
274 MemoryBuffer::getFileOrSTDIN(Path);
277 return !zlib::isAvailable() || CRCHash == zlib::crc32(MB.get()->getBuffer());
280 static bool findDebugBinary(const std::string &OrigPath,
281 const std::string &DebuglinkName, uint32_t CRCHash,
282 std::string &Result) {
283 std::string OrigRealPath = OrigPath;
284 #if defined(HAVE_REALPATH)
285 if (char *RP = realpath(OrigPath.c_str(), nullptr)) {
290 SmallString<16> OrigDir(OrigRealPath);
291 llvm::sys::path::remove_filename(OrigDir);
292 SmallString<16> DebugPath = OrigDir;
293 // Try /path/to/original_binary/debuglink_name
294 llvm::sys::path::append(DebugPath, DebuglinkName);
295 if (checkFileCRC(DebugPath, CRCHash)) {
296 Result = DebugPath.str();
299 // Try /path/to/original_binary/.debug/debuglink_name
300 DebugPath = OrigRealPath;
301 llvm::sys::path::append(DebugPath, ".debug", DebuglinkName);
302 if (checkFileCRC(DebugPath, CRCHash)) {
303 Result = DebugPath.str();
306 // Try /usr/lib/debug/path/to/original_binary/debuglink_name
307 DebugPath = "/usr/lib/debug";
308 llvm::sys::path::append(DebugPath, llvm::sys::path::relative_path(OrigDir),
310 if (checkFileCRC(DebugPath, CRCHash)) {
311 Result = DebugPath.str();
317 static bool getGNUDebuglinkContents(const ObjectFile *Obj, std::string &DebugName,
321 for (const SectionRef &Section : Obj->sections()) {
323 Section.getName(Name);
324 Name = Name.substr(Name.find_first_not_of("._"));
325 if (Name == "gnu_debuglink") {
327 Section.getContents(Data);
328 DataExtractor DE(Data, Obj->isLittleEndian(), 0);
330 if (const char *DebugNameStr = DE.getCStr(&Offset)) {
331 // 4-byte align the offset.
332 Offset = (Offset + 3) & ~0x3;
333 if (DE.isValidOffsetForDataOfSize(Offset, 4)) {
334 DebugName = DebugNameStr;
335 CRCHash = DE.getU32(&Offset);
346 bool darwinDsymMatchesBinary(const MachOObjectFile *DbgObj,
347 const MachOObjectFile *Obj) {
348 ArrayRef<uint8_t> dbg_uuid = DbgObj->getUuid();
349 ArrayRef<uint8_t> bin_uuid = Obj->getUuid();
350 if (dbg_uuid.empty() || bin_uuid.empty())
352 return !memcmp(dbg_uuid.data(), bin_uuid.data(), dbg_uuid.size());
355 ObjectFile *LLVMSymbolizer::lookUpDsymFile(const std::string &ExePath,
356 const MachOObjectFile *MachExeObj, const std::string &ArchName) {
357 // On Darwin we may find DWARF in separate object file in
358 // resource directory.
359 std::vector<std::string> DsymPaths;
360 StringRef Filename = sys::path::filename(ExePath);
361 DsymPaths.push_back(getDarwinDWARFResourceForPath(ExePath, Filename));
362 for (const auto &Path : Opts.DsymHints) {
363 DsymPaths.push_back(getDarwinDWARFResourceForPath(Path, Filename));
365 for (const auto &path : DsymPaths) {
366 ErrorOr<OwningBinary<Binary>> BinaryOrErr = createBinary(path);
367 std::error_code EC = BinaryOrErr.getError();
368 if (EC != errc::no_such_file_or_directory && !error(EC)) {
369 OwningBinary<Binary> B = std::move(BinaryOrErr.get());
371 getObjectFileFromBinary(B.getBinary(), ArchName);
372 const MachOObjectFile *MachDbgObj =
373 dyn_cast<const MachOObjectFile>(DbgObj);
374 if (!MachDbgObj) continue;
375 if (darwinDsymMatchesBinary(MachDbgObj, MachExeObj)) {
376 addOwningBinary(std::move(B));
384 LLVMSymbolizer::ObjectPair
385 LLVMSymbolizer::getOrCreateObjects(const std::string &Path,
386 const std::string &ArchName) {
387 const auto &I = ObjectPairForPathArch.find(std::make_pair(Path, ArchName));
388 if (I != ObjectPairForPathArch.end())
390 ObjectFile *Obj = nullptr;
391 ObjectFile *DbgObj = nullptr;
392 ErrorOr<OwningBinary<Binary>> BinaryOrErr = createBinary(Path);
393 if (!error(BinaryOrErr.getError())) {
394 OwningBinary<Binary> &B = BinaryOrErr.get();
395 Obj = getObjectFileFromBinary(B.getBinary(), ArchName);
397 ObjectPair Res = std::make_pair(nullptr, nullptr);
398 ObjectPairForPathArch[std::make_pair(Path, ArchName)] = Res;
401 addOwningBinary(std::move(B));
402 if (auto MachObj = dyn_cast<const MachOObjectFile>(Obj))
403 DbgObj = lookUpDsymFile(Path, MachObj, ArchName);
404 // Try to locate the debug binary using .gnu_debuglink section.
406 std::string DebuglinkName;
408 std::string DebugBinaryPath;
409 if (getGNUDebuglinkContents(Obj, DebuglinkName, CRCHash) &&
410 findDebugBinary(Path, DebuglinkName, CRCHash, DebugBinaryPath)) {
411 BinaryOrErr = createBinary(DebugBinaryPath);
412 if (!error(BinaryOrErr.getError())) {
413 OwningBinary<Binary> B = std::move(BinaryOrErr.get());
414 DbgObj = getObjectFileFromBinary(B.getBinary(), ArchName);
415 addOwningBinary(std::move(B));
422 ObjectPair Res = std::make_pair(Obj, DbgObj);
423 ObjectPairForPathArch[std::make_pair(Path, ArchName)] = Res;
428 LLVMSymbolizer::getObjectFileFromBinary(Binary *Bin,
429 const std::string &ArchName) {
432 ObjectFile *Res = nullptr;
433 if (MachOUniversalBinary *UB = dyn_cast<MachOUniversalBinary>(Bin)) {
434 const auto &I = ObjectFileForArch.find(
435 std::make_pair(UB, ArchName));
436 if (I != ObjectFileForArch.end())
438 ErrorOr<std::unique_ptr<ObjectFile>> ParsedObj =
439 UB->getObjectForArch(Triple(ArchName).getArch());
441 Res = ParsedObj.get().get();
442 ParsedBinariesAndObjects.push_back(std::move(ParsedObj.get()));
444 ObjectFileForArch[std::make_pair(UB, ArchName)] = Res;
445 } else if (Bin->isObject()) {
446 Res = cast<ObjectFile>(Bin);
452 LLVMSymbolizer::getOrCreateModuleInfo(const std::string &ModuleName) {
453 const auto &I = Modules.find(ModuleName);
454 if (I != Modules.end())
456 std::string BinaryName = ModuleName;
457 std::string ArchName = Opts.DefaultArch;
458 size_t ColonPos = ModuleName.find_last_of(':');
459 // Verify that substring after colon form a valid arch name.
460 if (ColonPos != std::string::npos) {
461 std::string ArchStr = ModuleName.substr(ColonPos + 1);
462 if (Triple(ArchStr).getArch() != Triple::UnknownArch) {
463 BinaryName = ModuleName.substr(0, ColonPos);
467 ObjectPair Objects = getOrCreateObjects(BinaryName, ArchName);
469 if (!Objects.first) {
470 // Failed to find valid object file.
471 Modules.insert(make_pair(ModuleName, (ModuleInfo *)nullptr));
474 DIContext *Context = nullptr;
475 if (auto CoffObject = dyn_cast<COFFObjectFile>(Objects.first)) {
476 // If this is a COFF object, assume it contains PDB debug information. If
477 // we don't find any we will fall back to the DWARF case.
478 std::unique_ptr<IPDBSession> Session;
479 PDB_ErrorCode Error = loadDataForEXE(PDB_ReaderType::DIA,
480 Objects.first->getFileName(), Session);
481 if (Error == PDB_ErrorCode::Success) {
482 Context = new PDBContext(*CoffObject, std::move(Session),
483 Opts.RelativeAddresses);
487 Context = new DWARFContextInMemory(*Objects.second);
489 ModuleInfo *Info = new ModuleInfo(Objects.first, Context);
490 Modules.insert(make_pair(ModuleName, Info));
494 std::string LLVMSymbolizer::printDILineInfo(DILineInfo LineInfo) const {
495 // By default, DILineInfo contains "<invalid>" for function/filename it
496 // cannot fetch. We replace it to "??" to make our output closer to addr2line.
497 static const std::string kDILineInfoBadString = "<invalid>";
498 std::stringstream Result;
499 if (Opts.PrintFunctions != FunctionNameKind::None) {
500 std::string FunctionName = LineInfo.FunctionName;
501 if (FunctionName == kDILineInfoBadString)
502 FunctionName = kBadString;
503 else if (Opts.Demangle)
504 FunctionName = DemangleName(FunctionName);
505 Result << FunctionName << "\n";
507 std::string Filename = LineInfo.FileName;
508 if (Filename == kDILineInfoBadString)
509 Filename = kBadString;
510 Result << Filename << ":" << LineInfo.Line << ":" << LineInfo.Column << "\n";
514 #if !defined(_MSC_VER)
515 // Assume that __cxa_demangle is provided by libcxxabi (except for Windows).
516 extern "C" char *__cxa_demangle(const char *mangled_name, char *output_buffer,
517 size_t *length, int *status);
520 std::string LLVMSymbolizer::DemangleName(const std::string &Name) {
521 #if !defined(_MSC_VER)
522 // We can spoil names of symbols with C linkage, so use an heuristic
523 // approach to check if the name should be demangled.
524 if (Name.substr(0, 2) != "_Z")
527 char *DemangledName = __cxa_demangle(Name.c_str(), nullptr, nullptr, &status);
530 std::string Result = DemangledName;
534 char DemangledName[1024] = {0};
535 DWORD result = ::UnDecorateSymbolName(
536 Name.c_str(), DemangledName, 1023,
537 UNDNAME_NO_ACCESS_SPECIFIERS | // Strip public, private, protected
538 UNDNAME_NO_ALLOCATION_LANGUAGE | // Strip __thiscall, __stdcall, etc
539 UNDNAME_NO_THROW_SIGNATURES | // Strip throw() specifications
540 UNDNAME_NO_MEMBER_TYPE | // Strip virtual, static, etc specifiers
541 UNDNAME_NO_MS_KEYWORDS | // Strip all MS extension keywords
542 UNDNAME_NO_FUNCTION_RETURNS); // Strip function return types
544 return (result == 0) ? Name : std::string(DemangledName);
548 } // namespace symbolize