Error R6016 during process termination

Question

I have a non sistematic crash (R6016) when our application is about to terminate. The application is composed of several DLL.

After a little investigation, we found the problem inside a function of mscvrt ( VS 2010) :

_ptiddata __cdecl _getptd_noexit (
        void
        )
{
    _ptiddata ptd;
    DWORD   TL_LastError;

    TL_LastError = GetLastError();

#ifdef _M_IX86

    /*
     * Initialize FlsGetValue function pointer in TLS by calling __set_flsgetvalue()
     */

    if ( (ptd = (__set_flsgetvalue())(__flsindex)) == NULL ) {
#else  /* _M_IX86 */
    if ( (ptd = FLS_GETVALUE(__flsindex)) == NULL ) {
#endif  /* _M_IX86 */
        /*
         * no per-thread data structure for this thread. try to create
         * one.
         */
#ifdef _DEBUG
        extern void * __cdecl _calloc_dbg_impl(size_t, size_t, int, const char *, int, int *);
        if ((ptd = _calloc_dbg_impl(1, sizeof(struct _tiddata), _CRT_BLOCK, __FILE__, __LINE__, NULL)) != NULL) {
#else  /* _DEBUG */
        if ((ptd = _calloc_crt(1, sizeof(struct _tiddata))) != NULL) {
#endif  /* _DEBUG */

            if (FLS_SETVALUE(__flsindex, (LPVOID)ptd) ) {

                /*
                 * Initialize of per-thread data
                 */

                _initptd(ptd,NULL);

                ptd->_tid = GetCurrentThreadId();
                ptd->_thandle = (uintptr_t)(-1);
            }
            else {

                /*
                 * Return NULL to indicate failure
                 */

                _free_crt(ptd);
                ptd = NULL;
            }
        }
    }

    SetLastError(TL_LastError);

    return(ptd);
}

This function is called normally during process shutdown and normally it works. When it doesn't work it returns NULL; analyzing the code the point of failure are:

• calloc_crt returns NULL, ie it can't alloc memory
• FLS_SETVALUE returns 0

FLS_SETVALUE is a macro that calls FlsSetValue. It receives a global var, __flsindex, that is the index for fiber allocated with FlsAlloc. __flsindex is resetted at termination of process in the ctr function of _mtterm

I think that crt is shutting down prematurely, before another DLL that use it but i don't know why. Our program doesn't call CreateThread.

So my question is:

• how dll are terminated and in which order? is it deterministic or changes from execution to execution?
• how can I investigate further?