semaphore problems in NGPT 1.2.1, and some questions

Combining libgc and NGPT is a challenge too large for me, as I don't
understand the use of the stacks by both libgc and NGPT. But it led to
some interesting observations.

For example, I wonder what happened to the comments Nicholas Black
(Mar 27) en I ( Apr 7) made on pthreads-devel about referencing
unallocated memory in sem_init and sem_destroy? The problems persist
in 1.2.1.

Also, in pth_mutex_acquire(...), I noticed the lock acquired through
_pth_acquire_lock() even when tryonly is set. I thought
_pth_try_lock() should be anough in that case.

Furthermore, I wondered which functions are async-signal-safe?
sem_post() ought to be, but doesn't seem to. And e.g. Boehm's gc
(erroniously, IMHO) assumes pthread_self() is.

Bye,
        Marijn