Re: Detecting hard float/version in pkgsrc

[coypu%SDF.ORG@localhost]

On Sun, Jul 24, 2016 at 04:07:05PM +0200, Martin Husemann wrote:
> On Sun, Jul 24, 2016 at 01:37:53PM +0000, coypu%SDF.ORG@localhost wrote:
> > We could theoretically let users use ./bootstrap -native for those, put
> > in devel/cpuflags, too...
> 
> Devel/cpuflags is a good way around one part of this for locally build
> pkgs. It has the bonus that you can manually adjust the settings if your
> build machine is different.
> 
> It does not work well for officially provided binaries, where it will be
> hard to agree on a proper non-worst-case default. However, especially
> on ARM we do really want different binaries for various families, and
> I think we are trying to provide a good set of defaults for those already.
> 
> The other option is runtime selection, ideally based on sysctl variables
> we provide reflecting the available cpu features, e.g.
> 
> machdep.sse = 1
> machdep.sse2 = 1
> 
> on x86 or
> 
> machdep.fpu_present = 0
> machdep.hwdiv_present = 0
> machdep.neon_present = 0
> machdep.simd_present = 0
> machdep.simdex_present = 0
> 
> on arm.
> 
> This, of course, can only be patched per pkg and needs to be upstreamed.
> 
> Martin

I guess the case of Go is pretty silly. I'm not even sure whether
defining GOARM at build time has the effect I expect (which is: default
build optimizes for the given setup), and even if it does, the better
target for this request would be at Go itself.

Patches to avoid crashes in the absence of SSE2 are more common, but I
suspect nobody is in a rush to eliminate those to optimize for 686s or
amd64-on-i386.

Thanks.