Cpptrace
Cpptrace is a simple, portable, and self-contained C++ stacktrace library supporting C++11 and greater on Linux, macOS, and Windows including MinGW and Cygwin environments. The goal: Make stack traces simple for once.
Table of Contents
- 30-Second Overview
- FAQ
- In-Depth Documentation
- Supported Debug Formats
- Usage
- Platform Logistics
- Library Back-Ends
- Testing Methodology
- Notes About the Library and Future Work
- License
30-Second Overview
Generating stack traces is as easy as:
#include <cpptrace/cpptrace.hpp>
void trace() {
cpptrace::generate_trace().print();
}
Cpptrace can also retrieve function inlining information on optimized release builds:
Cpptrace provides access to resolved stack traces as well as lightweight raw traces (just addresses) that can be resolved later:
const auto raw_trace = cpptrace::generate_raw_trace();
// then later
raw_trace.resolve().print();
Cpptrace also provides exception types that store stack traces:
#include <cpptrace/cpptrace.hpp>
void trace() {
throw cpptrace::logic_error("This wasn't supposed to happen!");
}
Additional notable features:
- Utilities for demangling
- Utilities for catching
std::exception
s and wrapping them in traced exceptions - Signal-safe stack tracing
CMake FetchContent Usage
include(FetchContent)
FetchContent_Declare(
cpptrace
GIT_REPOSITORY https://github.com/jeremy-rifkin/cpptrace.git
GIT_TAG v0.3.1 # <HASH or TAG>
)
FetchContent_MakeAvailable(cpptrace)
target_link_libraries(your_target cpptrace::cpptrace)
# On windows copy cpptrace.dll to the same directory as the executable for your_target
if(WIN32)
add_custom_command(
TARGET your_target POST_BUILD
COMMAND ${CMAKE_COMMAND} -E copy_if_different
$<TARGET_FILE:cpptrace::cpptrace>
$<TARGET_FILE_DIR:your_target>
)
endif()
Be sure to configure with -DCMAKE_BUILD_TYPE=Debug
or -DDCMAKE_BUILD_TYPE=RelWithDebInfo
for symbols and line
information.
On macos a little extra work to generate a .dSYM file is required, see Platform Logistics below.
For other ways to use the library, such as through package managers, a system-wide installation, or on a platform without internet access see Usage below.
FAQ
<stacktrace>
?
What about C++23 Some day C++23's <stacktrace>
will be ubiquitous. And maybe one day the msvc implementation will be acceptable.
The original motivation for cpptrace was to support projects using older C++ standards and as the library has grown its
functionality has extended beyond the standard library's implementation.
Cpptrace also provides additional functionality including being able to
In-Depth Documentation
namespace cpptrace
cpptrace::generate_trace()
can be used to generate a stacktrace object at the current call site. Resolved frames can
be accessed from this object with .frames
and also the trace can be printed with .print()
. Cpptrace also provides a
method to get lightweight raw traces, which are just vectors of program counters, which can be resolved at a later time.
Note: Debug info (-g
//Z7
//Zi
//DEBUG
) is generally required for good trace information.
Note: Currently on Mac .dSYM files are required, which can be generated with dsymutil yourbinary
. A cmake snippet
for generating these is provided in Platform Logistics below.
All functions are thread-safe unless otherwise noted.
Stack Traces
The core resolved stack trace object. Generate a trace with cpptrace::generate_trace()
or
cpptrace::stacktrace::current()
. On top of a set of helper functions struct stacktrace
allows
direct access to frames as well as iterators.
namespace cpptrace {
// Some type sufficient for an instruction pointer, currently always an alias to std::uintptr_t
using frame_ptr = std::uintptr_t;
struct stacktrace_frame {
frame_ptr address;
// nullable<T> represents a nullable integer. More docs later.
nullable<std::uint32_t> line;
nullable<std::uint32_t> column;
std::string filename;
std::string symbol;
bool is_inline;
bool operator==(const stacktrace_frame& other) const;
bool operator!=(const stacktrace_frame& other) const;
std::string to_string() const;
/* operator<<(ostream, ..) and std::format support exist for this object */
};
struct stacktrace {
std::vector<stacktrace_frame> frames;
// here as a drop-in for std::stacktrace
static stacktrace current(std::size_t skip = 0);
static stacktrace current(std::size_t skip, std::size_t max_depth);
void print() const;
void print(std::ostream& stream) const;
void print(std::ostream& stream, bool color) const;
std::string to_string(bool color = false) const;
void clear();
bool empty() const noexcept;
/* operator<<(ostream, ..), std::format support, and iterators exist for this object */
};
stacktrace generate_trace(std::size_t skip = 0);
stacktrace generate_trace(std::size_t skip, std::size_t max_depth);
}
Object Traces
Object traces contain the most basic information needed to construct a stack trace outside the currently running executable. It contains the raw address, the address in the binary (ASLR and the object file's memory space and whatnot is resolved), and the path to the object the instruction pointer is located in.
namespace cpptrace {
struct object_frame {
std::string object_path;
frame_ptr raw_address;
frame_ptr object_address;
};
struct object_trace {
std::vector<object_frame> frames;
static object_trace current(std::size_t skip = 0);
static object_trace current(std::size_t skip, std::size_t max_depth);
stacktrace resolve() const;
void clear();
bool empty() const noexcept;
/* iterators exist for this object */
};
object_trace generate_object_trace(std::size_t skip = 0);
object_trace generate_object_trace(std::size_t skip, std::size_t max_depth);
}
Raw Traces
Raw trace access: A vector of program counters. These are ideal for traces you want to resolve later.
Note it is important executables and shared libraries in memory aren't somehow unmapped otherwise libdl calls (and
GetModuleFileName
in windows) will fail to figure out where the program counter corresponds to.
namespace cpptrace {
struct raw_trace {
std::vector<frame_ptr> frames;
static raw_trace current(std::size_t skip = 0);
static raw_trace current(std::size_t skip, std::size_t max_depth);
object_trace resolve_object_trace() const;
stacktrace resolve() const;
void clear();
bool empty() const noexcept;
/* iterators exist for this object */
};
raw_trace generate_raw_trace(std::size_t skip = 0);
raw_trace generate_raw_trace(std::size_t skip, std::size_t max_depth);
}
Utilities
cpptrace::demangle
provides a helper function for name demangling, since it has to implement that helper internally
anyways.
The library makes an attempt to fail silently and continue during trace generation if any errors are encountered.
cpptrace::absorb_trace_exceptions
can be used to configure whether these exceptions are absorbed silently internally
or wether they're rethrown to the caller.
cpptrace::experimental::set_cache_mode
can be used to control time-memory tradeoffs within the library. By default
speed is prioritized. If using this function, set the cache mode at the very start of your program before any traces are
performed.
cpptrace::isatty
and the fileno definitions are useful for deciding whether to use color when printing stack taces.
cpptrace::register_terminate_handler()
is a helper function to set a custom std::terminate
handler that prints a
stack trace from a cpptrace exception (more info below) and otherwise behaves like the normal terminate handler.
namespace cpptrace {
std::string demangle(const std::string& name);
void absorb_trace_exceptions(bool absorb);
bool isatty(int fd);
extern const int stdin_fileno;
extern const int stderr_fileno;
extern const int stdout_fileno;
void register_terminate_handler();
enum class cache_mode {
// Only minimal lookup tables
prioritize_memory,
// Build lookup tables but don't keep them around between trace calls
hybrid,
// Build lookup tables as needed
prioritize_speed
};
namespace experimental {
void set_cache_mode(cache_mode mode);
}
}
Traced Exceptions
Cpptrace provides an interface for a traced exceptions, cpptrace::exception
, as well as a set of exception classes
that that generate stack traces when thrown. These exceptions generate relatively lightweight raw traces and resolve
symbols and line numbers lazily if and when requested.
The basic interface is:
namespace cpptrace {
class exception : public std::exception {
public:
virtual const char* what() const noexcept = 0; // The what string both the message and trace
virtual const char* message() const noexcept = 0;
virtual const stacktrace& trace() const noexcept = 0;
};
}
There are two ways to go about traced exception objects: Traces can be resolved eagerly or lazily. Cpptrace provides the basic implementation of exceptions as lazy exceptions. I hate to have anything about the implementation exposed in the interface or type system but this seems to be the best way to do this.
namespace cpptrace {
class lazy_exception : public exception {
mutable detail::lazy_trace_holder trace_holder; // basically std::variant<raw_trace, stacktrace>, more docs later
mutable std::string what_string;
protected:
explicit lazy_exception(std::size_t skip, std::size_t max_depth) noexcept;
explicit lazy_exception(std::size_t skip) noexcept;
public:
explicit lazy_exception() noexcept : lazy_exception(1) {}
const char* what() const noexcept override;
const char* message() const noexcept override;
const stacktrace& trace() const noexcept override;
};
}
cpptrace::lazy_exception
can be freely thrown or overridden. Generally message()
is the only field to override.
Lastly cpptrace provides an exception class that takes a user-provided message, cpptrace::exception_with_message
, as
well as a number of traced exception classes resembling <stdexcept>
:
namespace cpptrace {
class CPPTRACE_EXPORT exception_with_message : public lazy_exception {
mutable std::string user_message;
protected:
explicit exception_with_message(std::string&& message_arg, std::size_t skip) noexcept;
explicit exception_with_message(std::string&& message_arg, std::size_t skip, std::size_t max_depth) noexcept;
public:
explicit exception_with_message(std::string&& message_arg) noexcept
: exception_with_message(std::move(message_arg), 1) {}
const char* message() const noexcept override;
};
// All stdexcept errors have analogs here. Same constructor as exception_with_message.
class logic_error : public exception_with_message { ... };
class domain_error : public exception_with_message { ... };
class invalid_argument : public exception_with_message { ... };
class length_error : public exception_with_message { ... };
class out_of_range : public exception_with_message { ... };
class runtime_error : public exception_with_message { ... };
class range_error : public exception_with_message { ... };
class overflow_error : public exception_with_message { ... };
class underflow_error : public exception_with_message { ... };
}
Wrapping std::exceptions
Cpptrace exceptions can provide great information for user-controlled exceptions. For non-cpptrace::exceptions that may originate outside of code you control, e.g. the standard library, cpptrace provides some wrapper utilities that can rethrow these exceptions nested in traced cpptrace exceptions. The trace won't be perfect, the trace will start where the rapper caught it, but these utilities can provide good diagnostic information. Unfortunately this is the best solution for this problem, as far as I know.
std::vector<int> foo = {1, 2, 3};
CPPTRACE_WRAP_BLOCK(
foo.at(4) = 2;
foo.at(5)++;
);
std::cout<<CPPTRACE_WRAP(foo.at(12))<<std::endl;
Exception handling with cpptrace
Working with cpptrace exceptions in your code:
try {
foo();
} catch(cpptrace::exception& e) {
// Prints the exception info and stack trace, conditionally enabling color codes depending on
// whether stderr is a terminal
std::cerr << "Error: " << e.message() << '\n';
e.trace().print(std::cerr, cpptrace::isatty(cpptrace::stderr_fileno));
} catch(std::exception& e) {
std::cerr << "Error: " << e.what() << '\n';
}
Additionally cpptrace provides a custom std::terminate
handler that prints a stack trace from a cpptrace exception and otherwise behaves like the normal terminate handler and prints the stack trace involved in reaching std::terminate
.
The stack trace to std::terminate
may be helpful or it may not, it depends on the implementation, but often if an
implementation can't find an appropriate catch
while unwinding it will jump directly to std::terminate
giving
good information.
To register this custom handler:
cpptrace::register_terminate_handler();
Signal-Safe Tracing
Signal-safe stack tracing is very useful for debugging application crashes, e.g. SIGSEGVs or SIGTRAPs, but it's very difficult to do correctly and most implementations I see online do this incorrectly.
In order to do this full process safely the way to go is collecting basic information in the signal handler and then either resolving later or handing that information to another process to resolve.
It's not as simple as calling cpptrace::generate_trace().print()
, though you might be able to get
away with that, but this is what is needed to really do this safely.
The safe API is as follows:
namespace cpptrace {
std::size_t safe_generate_raw_trace(frame_ptr* buffer, std::size_t size, std::size_t skip = 0);
std::size_t safe_generate_raw_trace(frame_ptr* buffer, std::size_t size, std::size_t skip, std::size_t max_depth);
struct safe_object_frame {
frame_ptr raw_address;
frame_ptr address_relative_to_object_start; // object base address must yet be added
char object_path[CPPTRACE_PATH_MAX + 1];
object_frame resolve() const; // To be called outside a signal handler. Not signal safe.
};
void get_safe_object_frame(frame_ptr address, safe_object_frame* out);
}
Note: Not all back-ends and platforms support these interfaces. If signal-safe unwinding isn't supported
safe_generate_raw_trace
will just produce an empty trace and if object information can't be resolved in a signal-safe
way then get_safe_object_frame
will not populate fields beyond the raw_address
.
Another big note: Calls to shared objects can be lazy-loaded where the first call to the shared object invokes
non-signal-safe functions such as malloc()
. To avoid this, call these routines in main()
ahead of a signal handler
to "warm up" the library.
Because signal-safe tracing is an involved process, I have written up a comprehensive overview of what is involved at signal-safe-tracing.md.
Utility Types
A couple utility types are used to provide the library with a good interface.
nullable<T>
is used for a nullable integer type. Internally the maximum value for T
is used as a
sentinel. std::optional
would be used if this library weren't c++11. But, nullable<T>
provides
an std::optional
-like interface and it's less heavy-duty for this use than an std::optional
.
detail::lazy_trace_holder
is a utility type for lazy_exception
used in place of an
std::variant<raw_trace, stacktrace>
.
namespace cpptrace {
template<typename T, typename std::enable_if<std::is_integral<T>::value, int>::type = 0>
struct nullable {
T raw_value;
nullable& operator=(T value)
bool has_value() const noexcept;
T& value() noexcept;
const T& value() const noexcept;
T value_or(T alternative) const noexcept;
void swap(nullable& other) noexcept;
void reset() noexcept;
bool operator==(const nullable& other) const noexcept;
bool operator!=(const nullable& other) const noexcept;
constexpr static nullable null() noexcept; // returns a null instance
};
namespace detail {
class lazy_trace_holder {
bool resolved;
union {
raw_trace trace;
stacktrace resolved_trace;
};
public:
// constructors
lazy_trace_holder() : trace() {}
explicit lazy_trace_holder(raw_trace&& _trace);
explicit lazy_trace_holder(stacktrace&& _resolved_trace);
// logistics
lazy_trace_holder(const lazy_trace_holder& other);
lazy_trace_holder(lazy_trace_holder&& other) noexcept;
lazy_trace_holder& operator=(const lazy_trace_holder& other);
lazy_trace_holder& operator=(lazy_trace_holder&& other) noexcept;
~lazy_trace_holder();
// access
stacktrace& get_resolved_trace();
const stacktrace& get_resolved_trace() const; // throws if not already resolved
private:
void clear();
};
}
}
Supported Debug Formats
Format | Supported |
---|---|
DWARF in binary | βοΈ |
DWARF in separate binary (binary gnu debug link) | οΈοΈβοΈ |
DWARF in separate binary (split dwarf) | βοΈ |
DWARF in dSYM | βοΈ |
DWARF in via Mach-O debug map | Soon |
Windows debug symbols in PDB | βοΈ |
DWARF5 added DWARF package files. As far as I can tell no compiler implements these yet.
Usage
CMake FetchContent
With CMake FetchContent:
include(FetchContent)
FetchContent_Declare(
cpptrace
GIT_REPOSITORY https://github.com/jeremy-rifkin/cpptrace.git
GIT_TAG v0.3.1 # <HASH or TAG>
)
FetchContent_MakeAvailable(cpptrace)
target_link_libraries(your_target cpptrace::cpptrace)
It's as easy as that. Cpptrace will automatically configure itself for your system. Note: On windows and macos some extra work is required, see Platform Logistics below.
Be sure to configure with -DCMAKE_BUILD_TYPE=Debug
or -DDCMAKE_BUILD_TYPE=RelWithDebInfo
for symbols and line
information.
System-Wide Installation
git clone https://github.com/jeremy-rifkin/cpptrace.git
git checkout v0.3.1
mkdir cpptrace/build
cd cpptrace/build
cmake .. -DCMAKE_BUILD_TYPE=Release
make -j
sudo make install
Using through cmake:
find_package(cpptrace REQUIRED)
target_link_libraries(<your target> cpptrace::cpptrace)
Be sure to configure with -DCMAKE_BUILD_TYPE=Debug
or -DDCMAKE_BUILD_TYPE=RelWithDebInfo
for symbols and line
information.
Or compile with -lcpptrace
:
g++ main.cpp -o main -g -Wall -lcpptrace
./main
If you get an error along the lines of
error while loading shared libraries: libcpptrace.so: cannot open shared object file: No such file or directory
You may have to run sudo /sbin/ldconfig
to create any necessary links and update caches so the system can find
libcpptrace.so (I had to do this on Ubuntu). Only when installing system-wide. Usually your package manger does this for
you when installing new libraries.
System-wide install on windows
git clone https://github.com/jeremy-rifkin/cpptrace.git
git checkout v0.3.1
mkdir cpptrace/build
cd cpptrace/build
cmake .. -DCMAKE_BUILD_TYPE=Release
msbuild cpptrace.sln
msbuild INSTALL.vcxproj
Note: You'll need to run as an administrator in a developer powershell, or use vcvarsall.bat distributed with visual studio to get the correct environment variables set.
Local User Installation
To install just for the local user (or any custom prefix):
git clone https://github.com/jeremy-rifkin/cpptrace.git
git checkout v0.3.1
mkdir cpptrace/build
cd cpptrace/build
cmake .. -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=$HOME/wherever
make -j
sudo make install
Using through cmake:
find_package(cpptrace REQUIRED PATHS $ENV{HOME}/wherever)
target_link_libraries(<your target> cpptrace::cpptrace)
Using manually:
g++ main.cpp -o main -g -Wall -I$HOME/wherever/include -L$HOME/wherever/lib -lcpptrace
Package Managers
Conan
Cpptrace is available through conan at https://conan.io/center/recipes/cpptrace.
[requires]
cpptrace/0.3.1
[generators]
CMakeDeps
CMakeToolchain
[layout]
cmake_layout
# ...
find_package(cpptrace REQUIRED)
# ...
target_link_libraries(YOUR_TARGET cpptrace::cpptrace)
Vcpkg
vcpkg install cpptrace
find_package(cpptrace CONFIG REQUIRED)
target_link_libraries(main PRIVATE cpptrace::cpptrace)
Platform Logistics
Windows and macos require a little extra work to get everything in the right place
Copying the library .dll on windows:
# Copy the cpptrace.dll on windows to the same directory as the executable for your_target.
# Not required if static linking.
if(WIN32)
add_custom_command(
TARGET your_target POST_BUILD
COMMAND ${CMAKE_COMMAND} -E copy_if_different
$<TARGET_FILE:cpptrace::cpptrace>
$<TARGET_FILE_DIR:your_target>
)
endif()
Generating a .dSYM file on macos:
In xcode cmake this can be done with
set_target_properties(your_target PROPERTIES XCODE_ATTRIBUTE_DEBUG_INFORMATION_FORMAT "dwarf-with-dsym")
And outside xcode this can be done with dsymutil yourbinary
:
# Create a .dSYM file on macos. Currently required, but hopefully not for long
if(APPLE)
add_custom_command(
TARGET your_target
POST_BUILD
COMMAND dsymutil $<TARGET_FILE:your_target>
)
endif()
Library Back-Ends
Cpptrace supports a number of back-ends to produce stack traces. Stack traces are produced in roughly three steps: Unwinding, symbol resolution, and demangling.
The library's CMake automatically configures itself for what your system supports. The ideal configuration is as follows:
Platform | Unwinding | Symbols | Demangling |
---|---|---|---|
Linux | _Unwind |
libdwarf | cxxabi.h |
MacOS | _Unwind |
libdwarf | cxxabi.h |
Windows | StackWalk64 |
dbghelp | No demangling needed |
MinGW | StackWalk64 |
libdwarf + dbghelp | cxxabi.h |
Support for these back-ends is the main development focus and they should work well. If you want to use a different back-end such as addr2line, for example, you can configure the library to do so.
Unwinding
Library | CMake config | Platforms | Info |
---|---|---|---|
libgcc unwind | CPPTRACE_UNWIND_WITH_UNWIND |
linux, macos, mingw | Frames are captured with libgcc's _Unwind_Backtrace , which currently produces the most accurate stack traces on gcc/clang/mingw. Libgcc is often linked by default, and llvm has something equivalent. |
execinfo.h | CPPTRACE_UNWIND_WITH_EXECINFO |
linux, macos | Frames are captured with execinfo.h 's backtrace , part of libc on linux/unix systems. |
winapi | CPPTRACE_UNWIND_WITH_WINAPI |
windows, mingw | Frames are captured with CaptureStackBackTrace . |
dbghelp | CPPTRACE_UNWIND_WITH_DBGHELP |
windows, mingw | Frames are captured with StackWalk64 . |
dbghelp | CPPTRACE_UNWIND_WITH_LIBUNWIND |
linux, macos, windows, mingw | Frames are captured with libunwind. Note: This is the only back-end that requires a library to be installed by the user, and a CMAKE_PREFIX_PATH may also be needed. |
N/A | CPPTRACE_UNWIND_WITH_NOTHING |
all | Unwinding is not done, stack traces will be empty. |
Some back-ends (execinfo and CaptureStackBackTrace
) require a fixed buffer has to be created to read addresses into
while unwinding. By default the buffer can hold addresses for 200 frames (beyond the skip
frames). This is
configurable with CPPTRACE_HARD_MAX_FRAMES
.
Symbol resolution
Library | CMake config | Platforms | Info |
---|---|---|---|
libdwarf | CPPTRACE_GET_SYMBOLS_WITH_LIBDWARF |
linux, macos, mingw | Libdwarf is the preferred method for symbol resolution for cpptrace. Cpptrace will get it via FetchContent or find_package depending on CPPTRACE_USE_EXTERNAL_LIBDWARF . |
dbghelp | CPPTRACE_GET_SYMBOLS_WITH_DBGHELP |
windows | Dbghelp.h is the preferred method for symbol resolution on windows under msvc/clang and is supported on all windows machines. |
libbacktrace | CPPTRACE_GET_SYMBOLS_WITH_LIBBACKTRACE |
linux, macos*, mingw* | Libbacktrace is already installed on most systems or available through the compiler directly. For clang you must specify the absolute path to backtrace.h using CPPTRACE_BACKTRACE_PATH . |
addr2line | CPPTRACE_GET_SYMBOLS_WITH_ADDR2LINE |
linux, macos, mingw | Symbols are resolved by invoking addr2line (or atos on mac) via fork() (on linux/unix, and popen under mingw). |
libdl | CPPTRACE_GET_SYMBOLS_WITH_LIBDL |
linux, macos | Libdl uses dynamic export information. Compiling with -rdynamic is needed for symbol information to be retrievable. Line numbers won't be retrievable. |
N/A | CPPTRACE_GET_SYMBOLS_WITH_NOTHING |
all | No attempt is made to resolve symbols. |
*: Requires installation
One back-end should be used. For MinGW CPPTRACE_GET_SYMBOLS_WITH_LIBDWARF
and CPPTRACE_GET_SYMBOLS_WITH_DBGHELP
can
be used in conjunction.
Note for addr2line: By default cmake will resolve an absolute path to addr2line to bake into the library. This path can
be configured with CPPTRACE_ADDR2LINE_PATH
, or CPPTRACE_ADDR2LINE_SEARCH_SYSTEM_PATH
can be used to have the library
search the system path for addr2line
at runtime. This is not the default to prevent against path injection attacks.
Demangling
Lastly, depending on other back-ends used a demangler back-end may be needed.
Library | CMake config | Platforms | Info |
---|---|---|---|
cxxabi.h | CPPTRACE_DEMANGLE_WITH_CXXABI |
Linux, macos, mingw | Should be available everywhere other than msvc. |
dbghelp.h | CPPTRACE_DEMANGLE_WITH_WINAPI |
Windows | Demangle with UnDecorateSymbolName . |
N/A | CPPTRACE_DEMANGLE_WITH_NOTHING |
all | Don't attempt to do anything beyond what the symbol resolution back-end does. |
More?
There are plenty more libraries that can be used for unwinding, parsing debug information, and demangling. In the future more back-ends can be added. Ideally this library can "just work" on systems, without additional installation work.
Summary of Library Configurations
Summary of all library configuration options:
Back-ends:
CPPTRACE_GET_SYMBOLS_WITH_LIBDWARF=On/Off
CPPTRACE_GET_SYMBOLS_WITH_DBGHELP=On/Off
CPPTRACE_GET_SYMBOLS_WITH_LIBBACKTRACE=On/Off
CPPTRACE_GET_SYMBOLS_WITH_ADDR2LINE=On/Off
CPPTRACE_GET_SYMBOLS_WITH_LIBDL=On/Off
CPPTRACE_GET_SYMBOLS_WITH_NOTHING=On/Off
CPPTRACE_UNWIND_WITH_UNWIND=On/Off
CPPTRACE_UNWIND_WITH_LIBUNWIND=On/Off
CPPTRACE_UNWIND_WITH_EXECINFO=On/Off
CPPTRACE_UNWIND_WITH_WINAPI=On/Off
CPPTRACE_UNWIND_WITH_DBGHELP=On/Off
CPPTRACE_UNWIND_WITH_NOTHING=On/Off
CPPTRACE_DEMANGLE_WITH_CXXABI=On/Off
CPPTRACE_DEMANGLE_WITH_WINAPI=On/Off
CPPTRACE_DEMANGLE_WITH_NOTHING=On/Off
Back-end configuration:
CPPTRACE_BACKTRACE_PATH=<string>
: Path to libbacktrace backtrace.h, needed when compiling with clang/CPPTRACE_HARD_MAX_FRAMES=<number>
: Some back-ends write to a fixed-size buffer. This is the size of that buffer. Default is200
.CPPTRACE_ADDR2LINE_PATH=<string>
: Specify the absolute path to the addr2line binary for cpptrace to invoke. By default the config script will search for a binary and use that absolute path (this is to prevent against path injection).CPPTRACE_ADDR2LINE_SEARCH_SYSTEM_PATH=On/Off
: Specifies whether cpptrace should let the system search the PATH environment variable directories for the binary.
Other useful configurations:
CPPTRACE_BUILD_SHARED=On/Off
: Override forBUILD_SHARED_LIBS
.CPPTRACE_INCLUDES_WITH_SYSTEM=On/Off
: Marks cpptrace headers asSYSTEM
which will hide any warnings that aren't the fault of your project. Defaults to On.CPPTRACE_INSTALL_CMAKEDIR
: Override for the installation path for the cmake configs.CPPTRACE_USE_EXTERNAL_LIBDWARF=On/Off
: Get libdwarf fromfind_package
rather thanFetchContent
.
Testing:
CPPTRACE_BUILD_TESTING
Build small demo and test programCPPTRACE_BUILD_TEST_RDYNAMIC
Use-rdynamic
when compiling the test program
Testing Methodology
Cpptrace currently uses integration and functional testing, building and running under every combination of back-end
options. The implementation is based on github actions matrices and driven by python scripts located in the
ci/
folder. Testing used to be done by github actions matrices directly, however, launching hundreds of two
second jobs was extremely inefficient. Test outputs are compared against expected outputs located in
test/expected/
. Stack trace addresses may point to the address after an instruction depending on the
unwinding back-end, and the python script will check for an exact or near-match accordingly.
Notes About the Library and Future Work
For the most part I'm happy with the state of the library. But I'm sure that there is room for improvement and issues will exist. If you encounter any issue, please let me know! If you find any pain-points in the library, please let me know that too.
A note about performance: For handling of DWARF symbols there is a lot of room to explore for performance optimizations and time-memory tradeoffs. If you find the current implementation is either slow or using too much memory, I'd be happy to explore some of these options.
A couple things I'd like to improve in the future:
- On MacOS .dSYM files are required
- On Windows when collecting symbols with dbghelp (msvc/clang) parameter types are almost perfect but due to limitations in dbghelp the library cannot accurately show const and volatile qualifiers or rvalue references (these appear as pointers).
A couple features I'd like to add in the future:
Tracing from signal handlers- Tracing other thread's stacks
Showing inlined calls in the stack trace
License
This library is under the MIT license.
Cpptrace uses libdwarf on linux, macos, and mingw/cygwin unless configured to use something else. If this library is statically linked with libdwarf then the library's binary will itself be LGPL.