DevOps, CI/CD and Automation

I'm developing a huge multi-threaded application, in which many many threads access
a string object simultaneously. The program runs normally for few hours, but suddenly
it crashes with SEGV.

I've found that the reference count value of the string object is incorrect at the time of crash.
I investigated more and came to the fact that the guard mechanism for the reference count
differs between in the copy constructor and in the assignment operator of the string.
The former is instantiated in my application code and it uses mutex for guarding the reference
count, the latter is provided by libCstd_isa.so.1 and it uses InterlockedIncrement/Decrement.
Since InterlockedIncrement/Decrement can intrude into the mutex guard, the reference count
value becomes incorrect, I guess.

Some workarounds I thought of are:
1. use -xarch=v8plus option (the program above is compiled with no -x options)
But the application doesn't run on legacy sparc machine.
2. link libCstd.a statically
But the application size becomes large.
3. set LD_NOAUXFLTR environment variable at run time
But libc_psr.so.1 will also not be loaded.

Which (or any other) solution is recommended, or any patches out fix this problem?
I guess libCstd_isa.so.1 should provide the implementation of the copy constructor
using InterlockedIncrement.

Thanks.

OS: Solaris 8 10/00
Compiler: Forte 6 update 2
Patches: 111685-18, 108434-13, 109147-26