Problems throwing and catching exceptions on OS X with -fno-rtti
Asked Answered
V

3

20

The issue is somewhat similar to this question but the accepted answer does not really propose a solution or workaround.

In our project, we have a dylib and the main executalble. The dylib is compiled with -fno-rtti, while the executable does use RTTI. The problem happens when an exception (e.g. std::bad_alloc) is thrown from the dylib and is caught in the exe.

(Before you yell "exceptions need RTTI so you must have it on!", please note that the RTTI necessary for exceptions is always generated regardless of the -frtti or -fno-rtti setting. This is actually documented in the -fno-rtti flag description. The issue on OS X is that it's not generated in the same way)

After some investigation, the following was discovered:

  • In the dylib (-fno-rtti), there is a local copy of the exception's RTTI structures; in particular, the __ZTISt9bad_alloc symbol (typeinfo for std::bad_alloc).
  • The exe (-frtti) imports the typeinfo symbol from libstdc++.6.dylib and does not have a local copy

Since the exception handling code relies on comparing typeinfo pointers to determine exception match, the matching fails, and only the catch(...) succeeds.

So far I see the following options:

1) compile everything, or at least the files that throw and catch exceptions, with -frtti. This is doable but I don't like the idea of generating RTTI for everything even if we don't use it; and the list of files which work with exceptions is prone to get stale.

2) when linking the dylib, somehow make the linker throw away the weak exception definition from the object file and use the one from libstdc++.6.dylib. So far I was not successful.

3) ???

I made a small test illustrating the problem.

--- throw.cpp ---
#include <iostream>

#if defined(__GNUC__)
#define EXPORT __attribute__((visibility("default")))
#else
#define EXPORT __declspec(dllexport)
#endif

EXPORT void dothrow ()
{
   std::cout << "before throw" << std::endl;
   throw std::bad_alloc();
}

--- main.cpp ---
#include <stdio.h>
#include <iostream>

#if defined(__GNUC__)
#define IMPORT extern
#else
#define IMPORT __declspec(dllimport)
#endif

IMPORT void dothrow ();

int main (void) {
 try {
   std::cout << "trying lib->main exception" << std::endl;
   dothrow ();
 }
 catch ( const std::bad_alloc& )
 {
   std::cout << "caught bad_alloc in main - good." << std::endl;
 }
 catch (...)
 {
   std::cout << "caught ... in main - bad!" << std::endl;
 }
}

--- makefile ---
# for main exe
CFLAGS_RTTI=-m32 -frtti -fvisibility=hidden -fvisibility-inlines-hidden -shared-libgcc -funwind-tables
# for dylib
CFLAGS_NORTTI=-m32 -fno-rtti -fvisibility=hidden -fvisibility-inlines-hidden -shared-libgcc
# for linking; some switches which don't help
CLFLAGS=-Wl,-why_live,-warn_commons,-weak_reference_mismatches,error,-commons,error

all: test

test: libThrow.dylib main.o
    g++ $(CFLAGS_RTTI) -o test main.o -lthrow -L./ $(CLFLAGS)

main.o: main.cpp
    g++ $(CFLAGS_RTTI) -c -o main.o main.cpp

throw.o: throw.cpp
    g++ $(CFLAGS_NORTTI) -c -o throw.o throw.cpp

libThrow.dylib: throw.o
    g++ $(CFLAGS_NORTTI) -dynamiclib -o libThrow.dylib throw.o

clean:
    rm test main.o throw.o

Running:

$ ./test
trying lib->main exception
before throw
caught ... in main - bad!

Symbols of the files involved:

$ nm -m throw.o | grep bad_alloc
000001be (__TEXT,__textcoal_nt) weak private external __ZNSt9bad_allocC1Ev
000001be (__TEXT,__textcoal_nt) weak private external __ZNSt9bad_allocC1Ev
00000300 (__TEXT,__eh_frame) weak private external __ZNSt9bad_allocC1Ev.eh
         (undefined) external __ZNSt9bad_allocD1Ev
00000290 (__DATA,__const_coal) weak external __ZTISt9bad_alloc
000002a4 (__TEXT,__const_coal) weak external __ZTSSt9bad_alloc
         (undefined) external __ZTVSt9bad_alloc

$ nm -m libThrow.dylib | grep bad_alloc
00000ce6 (__TEXT,__text) non-external __ZNSt9bad_allocC1Ev
         (undefined) external __ZNSt9bad_allocD1Ev (from libstdc++)
00001050 (__DATA,__const) weak external __ZTISt9bad_alloc
00000e05 (__TEXT,__const) weak external __ZTSSt9bad_alloc
         (undefined) external __ZTVSt9bad_alloc (from libstdc++)

$ nm -m main.o | grep bad_alloc
         (undefined) external __ZTISt9bad_alloc

$ nm -m test | grep bad_alloc
         (undefined) external __ZTISt9bad_alloc (from libstdc++)

Note: similar compilation options on Linux and Windows works fine. I can throw exceptions from a shared object/dll and catch them in the main exe, even if they're compiled with different -frtti/-fno-rtti options.


EDIT: here's how I ended up solving it for the specific case of bad_alloc:

#if defined(__GLIBCXX__) || defined(_LIBCPP_VERSION)
 #define throw_nomem std::__throw_bad_alloc
#else
 #define throw_nomem throw std::bad_alloc
#endif

EXPORT void dothrow ()
{
   std::cout << "before throw" << std::endl;
   throw_nomem();
}

The function __throw_bad_alloc is imported from libstdc++.6.dylib and so always throws a correct type.

Valma answered 12/2, 2014 at 19:1 Comment(7)
(1) would seem at-least feasible if you had a way of squelching rtti for everything you don't need it for, such as a platform-dependant attribute, or turning it on in said-same manner. Interesting problem. I avoid throwing exceptions across dylibs for such reasons, but that may not be an option for you.Undulant
Could it be -fvisibility flags causing such behaviour?Classless
Can you even compile code that contains a throw statement with -fno-rtti? My GCC (admittedly not OSX, but still) doesn't let me do that at all (immediately bails out with an error), much to my grief. Also, it adds RTTI for every darn type in any translation unit whether the type may be thrown or not.Lombroso
@Damon: yes, it works for me in GCC 4.2 and 4.4 on Linux (don't have any newer ones to test right now).Valma
What I don't understand is why throw would seemingly use the weak __ZTISt9bad_alloc symbol in libThrow.dylib when there's a strong version of it to use (albeit indirect from libstdc++abi) in libstdc++.dylib which is probably being used by the catch. This seems like a "bug" (or at least a "quirk") with ld to me. Is that what seems to be going on?Fetlock
(fwiw, I tried messing with the order the libraries are linked -- which I have read can affect whether a strong or weak reference is used -- but it didn't seem to matter here.)Fetlock
@IgorSkochinsky Have you tried to compile with clang?Eurypterid
V
2

Well, even though I have accepted an answer it did not solve all problems. So I'm writing down the solution which did work in the end.

I made a small tool for post-processing the object files and marking the local symbols as UNDEF. This forces the linker to use definitions from libstdc++ and not local ones from the file. The basic approach of the tool is:

  1. load the Mach-O header
  2. walk the load commands and find the LC_SYMTAB command
  3. load the list of symbols (struct nlist) and the strings
  4. walk the symbols and look for those that we need (e.g. __ZTISt9bad_alloc)
  5. set the found symbols' type to N_UNDF|N_EXT.
  6. after processing, write the modified symbol table back to the file.

(I also made a similar implementation for ELF)

I post-process any file that's using std exceptions, either for throwing or for catching. To make sure the file list does not go stale, I added a post-link check for unwanted local symbols using nm.

This seems to resolve all the problems I've had so far.

Valma answered 17/3, 2014 at 16:2 Comment(0)
O
6

You can simply move all your "throw exceptions" infrastructure to the helper library with -frtti enabled - and link it to other stuff. Without actual code it's hard to tell, if this decomposition is possible or not.

Here is some sample code:

// Thrower.cc
void DoThrow() {
  throw std::bad_alloc;
}

// LibraryNoRTTI.cc
void f() {
  DoThrow();
}

// main.cc
int main() {
  try {
    f();
  }
  catch(std::bad_alloc&) {}
  return 0;
}

The simplest way is to move all your throw invocations to the separate functions with appropriate types, like: throw std::logical_error("message"); goes to void ThrowLogicError(const std::string& message) { ... }

If there is a problem with encapsulation (private exception classes), then you may make friends with throwing functions.


If you still want to use (throw/catch) exceptions inside the non-rtti library, then you have to make separation between internal exceptions and exceptions used in your library API.

The good way is to use native C++ throw-catch for internal purposes - and then rethrow some exceptions, using rtti-based library functions, to the outside - according to your interface:

// Thrower.cc
void Rethrow(const std::exception& e) {
  throw e;
}

// LibraryNoRTTI.cc
namespace {

void internal_stuff() {
  throw std::logical_error("something goes wrong!");
}

}  // namespace

// You even may explicitly specify the thrown exceptions in declaration:
void f() throw(std::logical_error) {
  try {
    internal_stuff();
  }
  catch(std::exception& e) {
    Rethrow(std::logical_error(std::string("Internal error: ") + e.what());
  }
}
Ovarian answered 3/3, 2014 at 6:2 Comment(2)
Thanks. It seems we'll have to go with this approach. I'm going to edit my post with a slightly more elegant solution for the specific case of bad_alloc.Valma
However, now we have the opposite problem: catch statements inside the dylib itself can't catch the exception :/Valma
A
3

Start Edit March 4, 2014
I think the Clang++ compiler has a better chance of obtaining the desired exception handling. I have found this Stack Overflow post: Clang and the default compiler in OS X Lion. The post has helpful script lines for modifying ~/.bashrc to change the system default compiler settings on Snow Leopard and how to use the LLVM GCC. For Clang, add inside the ~/.bashrc:

# Set Clang as the default compiler for the system
export CC=clang
export CFLAGS=-Qunused-arguments
export CPPFLAGS=-Qunused-arguments


If the c++ symbolic link is not present, either call clang++ directly or add the c++ link as desired (e.g.

ln -s /usr/bin/clang++ c++ 

). It is a good idea to check all symbolic links within the /usr/bin by running:

ls -l `which lynx` | more


On my Mavericks command line tools installation c++ points to clang++ and cc points to clang. The g++ compiler version says:

$ g++ --version
Configured with: --prefix=/Applications/Xcode.app/Contents/Developer/usr --with-gxx-   include-dir=/usr/include/c++/4.2.1
Apple LLVM version 5.0 (clang-500.2.79) (based on LLVM 3.3svn)
Target: x86_64-apple-darwin13.0.0
Thread model: posix


The clang++ complier version says:

$clang++ --version
Apple LLVM version 5.0 (clang-500.2.79) (based on LLVM 3.3svn)
Target: x86_64-apple-darwin13.0.0
Thread model: posix


Notice that the g++ include directory path is set to /usr/include/c++/4.2.1, probably not the include path needed to resolve the issue.

MacPorts: Hopefully the Answer for any OS X version
The best solution I can find to obtain any Clang++ compiler version for any OS X version is to use the open source tool called MacPorts. There is extensive documentation in the MacPorts Guide. The application is called port and can be installed from an OS X installation package or obtain the source code and compile locally. The following is from installing MacPorts onto Snow Leopard. The other OS X versions should be similar. After obtaining MacPorts for Snow Leopard, run the port search command to observe all the different clang related ports available. For example, it looks like this:

$port search clang 


The partial list of search results from Snow Leopard 10.6.8 is:

clang-2.9 @2.9_13 (lang)
    C, C++, Objective C and Objective C++ compiler

clang-3.0 @3.0_12 (lang)
    C, C++, Objective C and Objective C++ compiler

clang-3.1 @3.1_7 (lang)
    C, C++, Objective C and Objective C++ compiler

clang-3.2 @3.2_2 (lang)
    C, C++, Objective C and Objective C++ compiler

clang-3.3 @3.3_2 (lang)
    C, C++, Objective C and Objective C++ compiler

clang-3.4 @3.4 (lang)
    C, C++, Objective C and Objective C++ compiler

clang-3.5 @3.5-r202097 (lang)
    C, C++, Objective C and Objective C++ compiler

clang_select @0.1 (sysutils)
    common files for selecting default clang version


Then I successfully installed clang-3.3 with: sudo port install clang-3.3. After this completes, see the available versions by typing port select --list clang. Then run the

sudo port select --set clang mp-clang-3.3

or similar. When I execute clang++ --version it says (as expected):

clang version 3.3 (tags/RELEASE_33/final)
Target: x86_64-apple-darwin10.8.0
Thread model: posix


Same for when the clang --version command is executed (after closing and restarting the terminal):

clang version 3.3 (tags/RELEASE_33/final)
Target: x86_64-apple-darwin10.8.0
Thread model: posix

There are MacPorts installation packages for many OS X versions (e.g. Leopard, Snow Leopard, Lion, Mountain Lion, Mavericks, etc.). I did not go any further back than Leopard with my search. If using an OS X older than Leopard, please review the MacPorts site thoroughly.


If curious about details on where to find Xcode 4.2 (or used to be able to obtain it), I have found this post regarding obtaining Xcode 4.2 for Snow Leopard Xcode 4.2 download for Snow Leopard. Then these two additional ones: Can i use the latest features of C++11 in XCode 4 or OSX Lion? [duplicate] and Can I use C++11 with Xcode?. After trying a couple links to see if the 4.2 Xcode remains available for Snow Leopard, no joy.


More than likely the MacPorts libc++ installation will be necessary to have full C++11 support. To install the more recent version, execute sudo port install libcxx. The contents of the /usr/lib will be overwritten with the current C++11 libraries (as necessary per MacPorts Ticket #42385: libcxx/libcxxabi: OS update can render system unusable

If libc++ still appears to be lacking, try this: "libc++" C++ Standard Library. Then use this:

$ export TRIPLE=-apple-
$ export MACOSX_DEPLOYMENT_TARGET=10.6
$ ./buildit

from How to build libc++ with LLVM/Clang 3.3 on Mac OS X 10.6 "Snow Leopard".

On OS X Lion, Mountain Lion, and Mavericks, they all have recent independent command line tools downloads on the Apple Developer site. The Clang version may be older than what one needs, so be sure to confirm which C++11 features are needed when using the Developer site command line tools' Clang.

End Edit March 4, 2014

The above macro detection may need to change from __GNUC__ to __clang__ or __clang_version__. It all depends on what the predefined compiler macros are for each OS X compiler, and the best way to detect as needed here. The Stack Overflow answer at: What predefined macro can I use to detect clang? should be helpful in configuring the command line to obtain them (e.g. clang++ -dM -E -x c /dev/null).

I have noticed when running the preceding example command that there is a predefined macro called __ENVIRONMENT_MAC_OS_X_VERSION_MIN_REQUIRED__. On the Mavericks clang++, the macro value is 1090. It might be necessary to have a family of #ifdef logic to set the appropriate EXPORT macro for each OS X clang++ compiler.

Algebra answered 3/3, 2014 at 0:50 Comment(7)
Switching to clang++ (Apple clang version 3.0 (tags/Apple/clang-211.10.1) (based on LLVM 3.0svn)) does not work. And yes, it's not available on 10.6 so we can't use it anyway.Valma
I am not very confident there is a simple 10.6 solution. As I complete tasks today, I will run my Snow Leopard system and research this further. I want to find an answer that resolves it for all OS X versions. As I find a promising solution(s), I will update and/or post anew.Algebra
@IgorSkochinsky See above for the details on how to obtain any Clang version for any OS X version with MacPorts. Snow Leopard is used as the example.Algebra
Well, this is a nice guide, but as I mentioned, switching to clang does not solve the problem. It still behaves in the same way regarding -fno-rtti.Valma
@IgorSkochinsky Yeah, I have not taken the time yet to try a couple more things compilation flag and linking wise, since the post by abyss.7. The updated solution from abyss.7 does resolve the issue, right?Algebra
While testing a small Shape example project in C++11, CLang 3.3 on Snow Leopard 10.6.8, you may have compiler errors that look similar to, "no member named 'llrint' in the global namespace." If so, add the -U__STRICT_ANSI__option to the invocation of clang++. Apparently the 10.6 cmath.h has a problem.Algebra
When using NetBeans 7.3.1 on Snow Leopard, it is necessary to create a new tool collection to set the NetBeans | Preferences | Build Tools base directory to the new compiler. If this is not done, NetBeans will not find the 3.3 clang++ compiler (or whatever version is installed via MacPorts). Also be sure the new include path is at the top of the include list for both C and C++ file types. NetBeans may automatically add the correct ones, but please review before compiling.Algebra
V
2

Well, even though I have accepted an answer it did not solve all problems. So I'm writing down the solution which did work in the end.

I made a small tool for post-processing the object files and marking the local symbols as UNDEF. This forces the linker to use definitions from libstdc++ and not local ones from the file. The basic approach of the tool is:

  1. load the Mach-O header
  2. walk the load commands and find the LC_SYMTAB command
  3. load the list of symbols (struct nlist) and the strings
  4. walk the symbols and look for those that we need (e.g. __ZTISt9bad_alloc)
  5. set the found symbols' type to N_UNDF|N_EXT.
  6. after processing, write the modified symbol table back to the file.

(I also made a similar implementation for ELF)

I post-process any file that's using std exceptions, either for throwing or for catching. To make sure the file list does not go stale, I added a post-link check for unwanted local symbols using nm.

This seems to resolve all the problems I've had so far.

Valma answered 17/3, 2014 at 16:2 Comment(0)

© 2022 - 2024 — McMap. All rights reserved.