[Valgrind-users] Possible helgrind false positive in static global?

[André O. <off...@gm...>]

Hi all,

I'm struggling with a helgrind error that I don't understand. I've 
stripped the code to a minimal example that shows the issue and pasted 
the code below. As far as I can tell, the code should be thread safe, 
and I've looked at the gcc assembly output, and I can't see anything 
wrong with it so, so far I don't suspect it is a gcc error (but I'm not 
certain).

The helgrind error only appears when I compile with -O0, -O1 or -O2, not 
with -O3. Also, when I call "get_int_no_warning" instead of "get_int", 
no helgrind errors are reported.

Would anybody be able to tell me if this is a helgrind false positive?

I compile with gcc (Debian 8.3.0-2) 8.3.0 and am using valgrind-3.14.0 
from Debian testing.

(the weird method is used in a codebase to create an instance of an 
object that is never destructed but isn't reported as leaked memory 
either -- it's a hack around some static initialization order problems).

Source code:

#include <new>
#include <iostream>
#include <vector>
#include <thread>
#include <memory>

int* get_int()
{
   static std::aligned_storage_t<sizeof(int), alignof(int)> storage;

   // The following line is pointed out in a helgrind warning:
   // "Possible data race during read of size 8..."
   // When using -O2, this conflicts with a previous write of size 8, 
also at this line...
   static int* ptr = new (reinterpret_cast<int*>(&storage)) int();

   return ptr; // But with -O0, the previous line conflicts with this 
return statement.
}

int* get_int_no_warning()
{
   static int x = 0;
   return &x;
}

void test()
{
   volatile int* a = get_int();
}

int main()
{
   std::thread a(test), b(test);
   a.join();
   b.join();
}

Thanks,
André