Re: -march=native and build failures

On 26/08/2013 8:43 a.m., Alex Rousskov wrote:
> On 08/25/2013 02:16 PM, Kinkie wrote:
>
>> how about only enabling it on 32-bit builds and testing it out?
>> If Large Rock needs it, then removing it altoghether is not a viable option.
> To test out 64-bit atomics I was talking about, you may try to build the
> Collapsed Forwarding branch on lp:
>
> https://code.launchpad.net/~measurement-factory/squid/collapsed-fwd
>
>
> Building current trunk code would not test that aspect because the
> current code does not use 64-bit atomics AFAIK. IIRC, they were removed
> some time ago to avoid exactly these problems (it was possible to avoid
> them by using smaller atomics back then).

All -march=native does is set CPU option by-default. A configure option
to drop it is reasonable.

If we can isolate the option(s) required by these atomics they should go
in as a separate explicit flag test and set those options regardless of
-march=native. That way we both know that they are always present and
can isolate problems from the unrelated options -march=native causes to
be set.

Amos
Received on Mon Aug 26 2013 - 04:59:43 MDT