Should pthread_kill be marked __THROW?

[Roland McGrath]

> I agree.  Calling a signal handler synchronously is not a concurrency
> problem in the sense of synchronization with other threads.  The signal
> handler can be considered concurrency, but then we should rather tell
> the compiler that pthread_kill is like a raise, or not?

I first read "a raise" as "a call to the 'raise' function", but now I think
you actually meant "raising a C++ exception".  Is that right?

Currently pthread_kill, kill{,pg}, sigqueue, and raise are all annotated
in the same way (with our __THROW macro).  All of these have the same
specified requirement that (in some circumstances) the signal handler run
"inside" the function.  

So raise is marked as never doing a raise.  That is the problem being cited.