[ILUG] Test-and-set in C/C++

[Paul Jakma]

On Fri, 1 Oct 2004, Brian Foster wrote:

> yes.
> you cannot.
> not in _portable_ C anyways.
> there is no way, none, nothing, nil, nana, of writing
> any form of an atomic read-modify-write in de jour
> (ISO) standard conformant portable C (or C++, AFAIK).
>
> this always has to be done with machine-specific code,
> usually assembler, such as the hint/example someone
> posted previously --- which was specific to the x86
> (I think) _and_ GNU gcc(1).  that sort of specificity
> is very common.

Note though that some unix platforms /do/ provide C wrappers to 
atomic operations somewhere: man test_and_set on IRIX, man atomic_ops 
on Solaris and less /usr/include/asm/bitops.h on Linux.

Not portable though.

I'd have to wonder why Vincent wants atomic test and set though...

For signals, just increment a flag integer^Wsig_atomic_t and let 
something poll it, and/or catch the signal some other way (eg EINTR 
from select) - and let that sort things and do the /real/ work, where 
the full thread API is nicely available. Eg, something like:

 	sig_atomic_t globalflag;

sig handler:

 	if (globalflag < -1)
 	  /* over flow, set an overflow flag? */
 	else
 	 globalflag++

is probably sufficient for most purposes (on 2's complement machines, 
which is most common machines, no?). The non-signal-context 
poll/check of flag should be able to tell whether overflow occured, 
and it'll just have to have to deal with that, but unless you expect 
to handle massive amounts of signals (or check globalflag /rarely/), 
it's very unlikely ever to.

regards,
-- 
Paul Jakma	paul at clubi.ie	paul at jakma.org	Key ID: 64A2FF6A
Fortune:
In computing, the mean time to failure keeps getting shorter.