bug#54635: [PATCH v2 5/5] gnu: Add wfmash.

[Efraim Flashner <efraim@flashner.co.il>]

[-- Attachment #1: Type: text/plain, Size: 1862 bytes --]

On Thu, Mar 31, 2022 at 03:07:49PM +0200, Maxime Devos wrote:
> Efraim Flashner schreef op do 31-03-2022 om 15:18 [+0300]:
> > > Arun Isaac schreef op do 31-03-2022 om 12:58 [+0530]:
> > > > +                       (("-march=native") ""))
> > > 
> > > This is also wrong for x86 systems because it makes the build non-
> > > reproducible.  Also, has upstream been informed about some of the
> > > compiler flags being architecture-specific?
> > 
> > I'm pretty sure upstream is aware of it, and the -mcx16 flag. That
> > whole phase doesn't need to be non-x86_64 only, upstream prefers it
> > that way to get fater results
> 
> wfmash could be written to detect CPU features at runtime and there is
> also --tune.  Also, upstream preferring march=native does not make the
> build reproducible.
> 
> > but IMO it would be fine to move it into a snippet.
> 
> It does not have to be in a snippet, it just needs to be reproducible
> (so no march=native, whether on x86 or not).

I suppose not, but the -mcx16 should be in the snippet, since it adds
compiler flags which only work on some architectures. And while we're at
it we can do the -march=native one too.

> Upstream seems to be aware of the non-x86
> (https://github.com/ekg/wfmash/issues/125) but they do not seem to be
> aware of the problems with march=native.
> 

I've pushed these patches with a few tweaks and an additional patch to
run a test suite based on the github workflow in the repository. The
entire check phase takes ~2 minutes on my pinebook pro, so aarch64
doesn't seem to need to skip some tests like riscv64 does.


-- 
Efraim Flashner   <efraim@flashner.co.il>   אפרים פלשנר
GPG key = A28B F40C 3E55 1372 662D  14F7 41AA E7DC CA3D 8351
Confidentiality cannot be guaranteed on emails sent or received unencrypted

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 833 bytes --]