The problem seemed to be a pipelining problem of write and read instructions when swapping the context. An isync instruction when returning from a context switch solves the potential pipelining problem.
Reason for the problem was that tast_exit function in thread_arch.c tried to release the thread a second time although it was already released in sched_task_exit. A simple check whether the thread is already released (sched_active_thread == NULL) solved the problem.
Xtensa newlib version requires pthread_setcancelstate as symbol. Therefore, the module pthread was always used, which in turn requires the module xtimer. The xtimer module, however, uses TIMER_DEV(0). Therefore, tests/timers failed for TIMER_DEV(0).
