Can I use Thread Sanitizer for OpenMP programs?

Question

Consider the following example:

#include <iostream> 

int main () {
    int i = 0;
    #pragma omp parallel
    {
        #pragma omp critical
        {
            ++i;
        }
    }
    std::cout << i;
}

Compiling with g++ -fopenmp -fsanitize=thread and running yields

WARNING: ThreadSanitizer: data race (pid=9576)
Read of size 4 at 0x7ffdc170f600 by thread T1:
#0 main._omp_fn.0 (a.out+0x000000400d20)
#1 gomp_thread_start /build/gcc/src/gcc-5.2.0/libgomp/team.c:118 (libgomp.so.1+0x00000000f42d)

Previous write of size 4 at 0x7ffdc170f600 by thread T2:
#0 main._omp_fn.0 (a.out+0x000000400d35)
#1 gomp_thread_start /build/gcc/src/gcc-5.2.0/libgomp/team.c:118 (libgomp.so.1+0x00000000f42d)

Location is stack of main thread.

Thread T1 (tid=9578, running) created by main thread at:
#0 pthread_create /build/gcc/src/gcc-5.2.0/libsanitizer/tsan/tsan_interceptors.cc:895 (libtsan.so.0+0x000000027a37)
#1 gomp_team_start /build/gcc/src/gcc-5.2.0/libgomp/team.c:796 (libgomp.so.1+0x00000000f98f)
#2 __libc_start_main (libc.so.6+0x00000002060f)

Thread T2 (tid=9579, running) created by main thread at:
#0 pthread_create /build/gcc/src/gcc-5.2.0/libsanitizer/tsan/tsan_interceptors.cc:895 (libtsan.so.0+0x000000027a37)
#1 gomp_team_start /build/gcc/src/gcc-5.2.0/libgomp/team.c:796 (libgomp.so.1+0x00000000f98f)
#2 __libc_start_main (libc.so.6+0x00000002060f)

SUMMARY: ThreadSanitizer: data race ??:0 main._omp_fn.0

As far as I can see, this is a false positive. Is there a way to avoid this?

(Something working with clang and libomp would be fine too.)

You would need to build the OpenMP runtime library with TSAN as well if you want to analyze the code; otherwise add a suppression? — Kerrek SB, Oct 08 '15 at 00:59
@KerrekSB I was afraid that this is the only answer. Building modern stuff on my Ubuntu 12.04 work machine and getting it to work is not very fun. — Baum mit Augen, Oct 08 '15 at 01:06
@KerrekSB Can you elaborate on the suppression stuff? Do I need to do that for every false hit? In our production code, there are many, and if I have to check all of them "by hand", the sanitizer kind of loses its sense. — Baum mit Augen, Oct 08 '15 at 01:15
I think you should be able to suppress by code location. I've never done that (the last suppression set I wrote was for Valgrind), but check out the wiki: https://github.com/google/sanitizers/wiki/ThreadSanitizerSuppressions — Kerrek SB, Oct 08 '15 at 08:52

score 15 · Accepted Answer · answered Mar 13 '19 at 16:57

15

Yes, at least with Clang this is relatively easy. You'll need to build libomp (which Clang uses instead of libgomp) with ThreadSanitizer support. This doesn't take that long:

git clone https://github.com/llvm/llvm-project
cd llvm-project
mkdir build
cd build
cmake -DLIBOMP_TSAN_SUPPORT=1 ../openmp
sudo cmake --build . --target install

(sudo and --target install is optional if you adjust the path to libomp.so below)

Now running your example works without any errors if you use this libomp.so instead of the system one:

clang++ -fsanitize=thread -fopenmp main.cpp
env LD_PRELOAD=/usr/local/lib/libomp.so ./a.out

answered Mar 13 '19 at 16:57

jhasse

2,379
1
30
40

Apparently this **only** works on Linux with a 64-bit Intel CPU. The CMake script checks `if("${CMAKE_SYSTEM_NAME}" MATCHES "Linux" AND ${INTEL64})`. – Cris Luengo May 27 '20 at 16:05
`INTEL64` will be set for AMD CPUs (or other x86_64 CPUs). See the root CMakeLists.txt of llvm-project – jhasse May 27 '20 at 17:10
Sure. But for me the important bit there was to boot Linux, this unfortunately doesn't work on macOS. I thought it was relevant information to add to the answer. Once I was on Linux, this worked brilliantly! Many thanks! – Cris Luengo May 27 '20 at 17:24

score 4 · Answer 2 · answered Feb 25 '16 at 06:14

4

even with the suppression you will still get false positives on the OpenMP runtime because there are some synchronization mechanism within the runtime that Tsan is not able to understand.

We worked on the OpenMP runtime to make Tsan understand this synchronization points and remove all the false positives.

Take a look to this project:

https://github.com/PRUNER/archer

Let me know if you need more help.

Best,

Simone

answered Feb 25 '16 at 06:14

Simone Atzeni

63
5

Thanks, I'll check it out when I get the chance. Might be a bit though, I'm working on something else now and a colleague does the sanitizer stuff. – Baum mit Augen Feb 25 '16 at 18:50

Can I use Thread Sanitizer for OpenMP programs?

2 Answers2

Linked