Seemingly incorrect behaviour with atomic ops on PowerPC

[James Molloy <jamesmol at transitive dot com>]

Hi,

I've noticed behaviour which doesn't match that described in the GCC 
manual for the __sync_bool_compare_and_swap builtin, when operating on 
objects of type 'bool'.

Take the following test program as an example, tested on a PowerPC ibook 
with GCC 4.3.0:

int main(int argc, char **argv)
{
 bool a = true;
 // Swap the value 'true' with 'false' and return true on success.
 bool b = __sync_bool_compare_and_swap(&a, true, false);

 return (b) 2 : 1;
}

This returns '1' - that is, the compare and swap failed - even there is 
no reason for it to fail. Interesting to note is that this behaviour is 
/not/ exhibited on x86, x86_64 or mips32 (the only other platforms I was 
able to test on), and neither is it exhibited if swapping false with 
true, for example:

int main(int argc, char **argv)
{
 bool a = false;
 // Swap the value 'false' with 'true' and return true on success.
 bool b = __sync_bool_compare_and_swap(&a, false, true);

 return (b) 2 : 1;
}

The above program returns '2'.

I'd be interested in any response - Am I right in saying that this is 
unintended behaviour or am I somehow using the function outside its 
intended remit?

Cheers,

James Molloy