Re: incorrect use of 'pure' attribute

Paul Eggert wrote:
> > Storing the same value in the same location (of
> > type 'int') is OK not only on those platforms where an 'int' access is atomic,
> > but even on architectures (such as m68k) where it isn't.
> 
> This variable is 'long', not 'int'.

Oops, right. But anyway, 'long' is the size of a machine word, and except
on m68k, for aligned accesses, loads and stores of such words are atomic.

> Come to think of it, though, on GNU/Hurd aren't we in a bit of trouble 
> here? One thread might be writing -1 while another is reading, and a 
> torn read might cause the latter to see "65535" instead of -1.

Indeed, I was only thinking about write/write races, but there are
also read/write races. The scenario you describe can only occur on
Hurd/m68k, a platform that will never exist.

> We needn't go to atomic accesses to fix this, I'd think, just to a 
> volatile bool saying whether the long int is initialized, along with 
> rejiggering the accesses to the long int.

Yes, adding 'volatile' would be an optimization: it would reduce the
probability that different threads execute the sysconf call again without
necessity.

Bruno