Re: autoconf stalls while checking "atomic builtins for 64 bit data types"

[Helge Deller <deller@gmx.de>]

Hi Christoph,

On 19.10.2017 00:01, Christoph Biedl wrote:
> John David Anglin wrote...
>
> > On 2017-10-18 5:04 PM, Helge Deller wrote:
> > > Dave submitted this patch to the parisc kernel mailing list, which most
> > > likely will fix the problem:
> > > https://patchwork.kernel.org/patch/9979587/
> > >
> > > Did anyone of you tested it?
>
> Wasn't aware of that, thanks for the pointer.
>
> > It's build tested but needs to be tested on a machine with a 32-bit kernel.
>
> Tested-by: Christoph Biedl <debian.axhn@manchmal.in-ulm.de>

Thanks for testing!
 
> Looks very good. On my system [PA8600 (PCX-W+) 64 bit CPU] with 32-bit
> kernel the reproducer exits immediatly, also rsyslog's configure now
> passes:
>
> | checking for __builtin_expect()... yes
> | checking whether the compiler provides atomic builtins... yes
> + checking whether the compiler provides atomic builtins for 64 bit data types... yes

Although the 32-bit parisc kernel now provides 64-bit atomics, I'd suggest
to stay on "native word width" when writing locking functions.
That means, on a 32-bit kernel ideally 32-bit atomics should be used.
We believe that our 64-bit atomic functions do work, but
there might be cornercases where it doesn't work 100% reliable.

> | checking semaphore.h usability... yes
>
> Let me know if I should check on 32-bit hardware, I might have access to
> one in three weeks.

Thanks, but that's not necessary.
It will work on PA32 then too.

Helge