Re: -march=native and build failures

On 08/25/2013 01:08 PM, Kinkie wrote:

> I've traced down at least some build failures to our use of -march=native.
> What apparently happens is that in some ocasions gcc emits some mmx
> instructions which fail with a SIGILL.
>
> We have a few choices:
> 1- leave the builds broken for now and try rework the code
> 2- add an --enable-extra-compiler-optimizations configure switch to
> optionally enable -march=native instead of unconditionally.
> 3- work around this on the broken nodes and on those only by using
> build trickery.
>
> Thoughts?
> I'd favor option 2'; we can then leave a specifically-broken build job
> so that when some change we make helps fix the build, we will know.

IIRC, -march=native or similar was needed to build Squid on some 32-bit
systems. Without it, the build (linking) failed due to the lack of
64-bit atomic operations if compiler assumes it needs to preserve
compatibility with much older 32 bit systems.

Current trunk does not use such operations, but the pending Large Rock
and Collapsed Forward code do (because swap_file_sz that they need to
keep in sync across Squid kids is 64 bits).

I hope this info helps you make the right decision but I do not know
enough about the related compiler options to pick the right choice among
the three offered.

Thank you,

Alex.
Received on Sun Aug 25 2013 - 20:05:11 MDT