Debugging Support
C++
debug
There are numerous things that can be done to improve the ease with
which C++ binaries are debugged when using the GNU tool chain. Here
are some of them.
Using g++
Compiler flags determine how debug information is transmitted
between compilation and debug or analysis tools.
The default optimizations and debug flags for a libstdc++ build
are -g -O2. However, both debug and optimization
flags can be varied to change debugging characteristics. For
instance, turning off all optimization via the -g -O0
-fno-inline flags will disable inlining and optimizations,
and add debugging information, so that stepping through all functions,
(including inlined constructors and destructors) is possible. In
addition, -fno-eliminate-unused-debug-types can be
used when additional debug information, such as nested class info,
is desired.
Or, the debug format that the compiler and debugger use to
communicate information about source constructs can be changed via
-gdwarf-2 or -gstabs flags: some debugging
formats permit more expressive type and scope information to be
shown in GDB. Expressiveness can be enhanced by flags like
-g3. The default debug information for a particular
platform can be identified via the value set by the
PREFERRED_DEBUGGING_TYPE macro in the gcc sources.
Many other options are available: please see "Options
for Debugging Your Program" in Using the GNU Compiler
Collection (GCC) for a complete list.
Debug Versions of Library Binary Files
If you would like debug symbols in libstdc++, there are two ways to
build libstdc++ with debug flags. The first is to run make from the
toplevel in a freshly-configured tree with
--enable-libstdcxx-debug
and perhaps
--enable-libstdcxx-debug-flags='...'
to create a separate debug build. Both the normal build and the
debug build will persist, without having to specify
CXXFLAGS, and the debug library will be installed in a
separate directory tree, in (prefix)/lib/debug. For
more information, look at the configuration section.
A second approach is to use the configuration flags
make CXXFLAGS='-g3 -fno-inline -O0' all
This quick and dirty approach is often sufficient for quick
debugging tasks, when you cannot or don't want to recompile your
application to use the debug mode.Memory Leak Hunting
There are various third party memory tracing and debug utilities
that can be used to provide detailed memory allocation information
about C++ code. An exhaustive list of tools is not going to be
attempted, but includes mtrace, valgrind,
mudflap, and the non-free commercial product
purify. In addition, libcwd has a
replacement for the global new and delete operators that can track
memory allocation and deallocation and provide useful memory
statistics.
Regardless of the memory debugging tool being used, there is one
thing of great importance to keep in mind when debugging C++ code
that uses new and delete: there are
different kinds of allocation schemes that can be used by
std::allocator . For implementation details, see the mt allocator documentation and
look specifically for GLIBCXX_FORCE_NEW.
In a nutshell, the default allocator used by
std::allocator is a high-performance pool allocator, and can
give the mistaken impression that in a suspect executable, memory is
being leaked, when in reality the memory "leak" is a pool being used
by the library's allocator and is reclaimed after program
termination.
For valgrind, there are some specific items to keep in mind. First
of all, use a version of valgrind that will work with current GNU
C++ tools: the first that can do this is valgrind 1.0.4, but later
versions should work at least as well. Second of all, use a
completely unoptimized build to avoid confusing valgrind. Third, use
GLIBCXX_FORCE_NEW to keep extraneous pool allocation noise from
cluttering debug information.
Fourth, it may be necessary to force deallocation in other libraries
as well, namely the "C" library. On linux, this can be accomplished
with the appropriate use of the __cxa_atexit or
atexit functions.
#include <cstdlib>
extern "C" void __libc_freeres(void);
void do_something() { }
int main()
{
atexit(__libc_freeres);
do_something();
return 0;
}
or, using __cxa_atexit:
extern "C" void __libc_freeres(void);
extern "C" int __cxa_atexit(void (*func) (void *), void *arg, void *d);
void do_something() { }
int main()
{
extern void* __dso_handle __attribute__ ((__weak__));
__cxa_atexit((void (*) (void *)) __libc_freeres, NULL,
&__dso_handle ? __dso_handle : NULL);
do_test();
return 0;
}
Suggested valgrind flags, given the suggestions above about setting
up the runtime environment, library, and test file, might be:
valgrind -v --num-callers=20 --leak-check=yes --leak-resolution=high --show-reachable=yes a.out
Data Race Hunting
All synchronization primitives used in the library internals need to be
understood by race detectors so that they do not produce false reports.
Two annotation macros are used to explain low-level synchronization
to race detectors:
_GLIBCXX_SYNCHRONIZATION_HAPPENS_BEFORE() and
_GLIBCXX_SYNCHRONIZATION_HAPPENS_AFTER().
By default, these macros are defined empty -- anyone who wants
to use a race detector needs to redefine them to call an
appropriate API.
Since these macros are empty by default when the library is built,
redefining them will only affect inline functions and template
instantiations which are compiled in user code. This allows annotation
of templates such as shared_ptr, but not code which is
only instantiated in the library. Code which is only instantiated in
the library needs to be recompiled with the annotation macros defined.
That can be done by rebuilding the entire
libstdc++.so file but a simpler
alternative exists for ELF platforms such as GNU/Linux, because ELF
symbol interposition allows symbols defined in the shared library to be
overridden by symbols with the same name that appear earlier in the
runtime search path. This means you only need to recompile the functions
that are affected by the annotation macros, which can be done by
recompiling individual files.
Annotating std::string and std::wstring
reference counting can be done by disabling extern templates (by defining
_GLIBCXX_EXTERN_TEMPLATE=-1) or by rebuilding the
src/string-inst.cc file.
Annotating the remaining atomic operations (at the time of writing these
are in ios_base::Init::~Init, locale::_Impl,
locale::facet and thread::_M_start_thread)
requires rebuilding the relevant source files.
The approach described above is known to work with the following race
detection tools:
DRD,
Helgrind, and
ThreadSanitizer.
With DRD, Helgrind and ThreadSanitizer you will need to define
the macros like this:
#define _GLIBCXX_SYNCHRONIZATION_HAPPENS_BEFORE(A) ANNOTATE_HAPPENS_BEFORE(A)
#define _GLIBCXX_SYNCHRONIZATION_HAPPENS_AFTER(A) ANNOTATE_HAPPENS_AFTER(A)
Refer to the documentation of each particular tool for details.
Using gdb
Many options are available for GDB itself: please see
"GDB features for C++" in the GDB documentation. Also
recommended: the other parts of this manual.
These settings can either be switched on in at the GDB command line,
or put into a .gdbint file to establish default debugging
characteristics, like so:
set print pretty on
set print object on
set print static-members on
set print vtbl on
set print demangle on
set demangle-style gnu-v3
Starting with version 7.0, GDB includes support for writing
pretty-printers in Python. Pretty printers for STL classes are
distributed with GCC from version 4.5.0. The most recent version of
these printers are always found in libstdc++ svn repository.
To enable these printers, check-out the latest printers to a local
directory:
svn co svn://gcc.gnu.org/svn/gcc/trunk/libstdc++-v3/python
Next, add the following section to your ~/.gdbinit The path must
match the location where the Python module above was checked-out.
So if checked out to: /home/maude/gdb_printers/, the path would be as
written in the example below.
python
import sys
sys.path.insert(0, '/home/maude/gdb_printers/python')
from libstdcxx.v6.printers import register_libstdcxx_printers
register_libstdcxx_printers (None)
end
The path should be the only element that needs to be adjusted in the
example. Once loaded, STL classes that the printers support
should print in a more human-readable format. To print the classes
in the old style, use the /r (raw) switch in the print command
(i.e., print /r foo). This will print the classes as if the Python
pretty-printers were not loaded.
For additional information on STL support and GDB please visit:
"GDB Support
for STL" in the GDB wiki. Additionally, in-depth
documentation and discussion of the pretty printing feature can be
found in "Pretty Printing" node in the GDB manual. You can find
on-line versions of the GDB user manual in GDB's homepage, at
"GDB: The GNU Project
Debugger" .
Tracking uncaught exceptions
The verbose
termination handler gives information about uncaught
exceptions which are killing the program. It is described in the
linked-to page.
Debug Mode The Debug Mode
has compile and run-time checks for many containers.
Compile Time Checking The Compile-Time
Checks Extension has compile-time checks for many algorithms.
Profile-based Performance Analysis The Profile-based
Performance Analysis Extension has performance checks for many
algorithms.