[cfe-users] Clang memory sanitizer: llvm-symbolizer problem

[Schlottke-Lakemper, Michael via cfe-users]

On 26 Nov 2015, at 17:43 , Kevin P. Fleming <kevin at kpfleming.us<mailto:kevin at kpfleming.us>> wrote:

Instead of putting the directory that contains libc++/libc++abi into LD_LIBRARY_PATH, use the system's 'normal' method of locating libraries, by ensuring that directory is included in the search path built up from /etc/ld.so.conf. If you do this, then an RPATH in an executable *will* take precedence over the default directory, because the default directory won't be specified in LD_LIBRARY_PATH (which has the highest precedence).
This, of course, would be the preferred solution. Unfortunately, while it is possible for me to influence the values in LD_LIBRARY_PATH, I cannot easily update the ldconfig cache on all cluster nodes (if at all). I am thus still interested in (and actively searching for) an alternative approach.

Michael


On Thu, Nov 26, 2015 at 4:09 AM, Schlottke-Lakemper, Michael via cfe-users <cfe-users at lists.llvm.org<mailto:cfe-users at lists.llvm.org>> wrote:

On 25 Nov 2015, at 20:13 , Alexey Samsonov <vonosmas at gmail.com<mailto:vonosmas at gmail.com>> wrote:

On Wed, Nov 25, 2015 at 4:12 AM, Schlottke-Lakemper, Michael <m.schlottke-lakemper at aia.rwth-aachen.de<mailto:m.schlottke-lakemper at aia.rwth-aachen.de>> wrote:

Thus I tried to compile a complete LLVM/Clang stack (with compiler-rt, libcxx, libcxxabi, libomp) using -DLLVM_USE_SANITIZER=MemoryWithOrigins. However, this did not work either, as apparently during the compilation process the memory sanitizer already comes to life and complains about use-of-unitialized values… Thus this approach seems to be a dead end.

Michael

On 24 Nov 2015, at 15:28 , Michael Schlottke-Lakemper <m.schlottke-lakemper at aia.rwth-aachen.de<mailto:m.schlottke-lakemper at aia.rwth-aachen.de>> wrote:

Hi folks,

When running our msan-instrumented simulation program, instead of a proper output I get the following error:

==12089==WARNING: MemorySanitizer: use-of-uninitialized-value
==12089==WARNING: Can't read from symbolizer at fd 14
/pds/opt/llvm/bin/llvm-symbolizer: symbol lookup error: /pds/opt/libcxx-20151121-r253770-clang-msan/lib/libc++abi.so.1: undefined symbol: __msan_origin_tls
==12089==WARNING: external symbolizer didn't start up correctly!

My setup for using the memory sanitizer is as follows:
- compile libc++/libc++abi with -fsanitize=memory
- compile test program with normal (=uninstrumented) LLVM/Clang installation using -fsanitize=memory
- put instrumented libcxx/libcxxabi library dirs into LD_LIBRARY_PATH
Can you pass -Wl,-rpath when you link your executable, to specify the path to instrumented libc++/libc++abi?

Yes, and if I do that together with setting LD_LIBRARY_PATH=“”, it works, thanks.

Why do you need to clear LD_LIBRARY_PATH to run MSan?
What libraries from LLVM lib dir harm running the MSan-ified executalbe? I thought that if you set rpath to correct path with instrumented libraries it should be enough…

Right now I have the problem that by default, LD_LIBRARY_PATH includes a directory with libc++/libc++abi in it. However, when running msan-instrumented executables to test them, the executable should pick up msan-instrumented versions of libc++/libc++abi while the subsequently used llvm-symbolizer needs to pick up the uninstrumented versions (otherwise I get the above-mentioned errors). Setting rpath is not enough, since the presence of libc++/libc++abi in LD_LIBRARY_PATH will override it. Compiling llvm-symbolizer in a way that it is able to run with msan-instrumented libc++ might be an option, but I don’t know how to do that (other than compiling all of LLVM with msan, which does not work either).

One possible workaround that I have been using is to create a shell script, call it “llvm-symbolizer”, put

LD_LIBRARY_PATH="$LLVM_DIR/lib:$LD_LIBRARY_PATH" $LLVM_DIR/bin/llvm-symbolizer $*

in it, and make sure that it is on the PATH before the actual llvm-symbolizer executable (which is located in LLVM_DIR/bin). However, this is clearly a workaround and I am looking for a more robust and more “canonical” solution.

Any suggestions?

Michael


_______________________________________________
cfe-users mailing list
cfe-users at lists.llvm.org<mailto:cfe-users at lists.llvm.org>
http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-users



-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-users/attachments/20151130/0a29de11/attachment.html>