Le 06/11/2016 à 11:59, Emilio Pozuelo Monfort a écrit :
Hi Sylvestre,
On Fri, 4 Nov 2016 09:50:27 +0100 Sylvestre Ledru <[email protected]> wrote:
severity 842143 normal
thanks
See the bug report.
If I understood things correctly, the problem is that by default, LLVM assumes
NEON extensions are available when compiling for armhf. That causes rdeps to
fail with illegal instruction when running on armhf hardware without NEON
support, however rare that may be, which is officially supported on armhf.
According to the upstream bug, this can be changed by passing a flag to the
compiler. That's not good enough as that means we have to change all rdeps.
Instead, the default should be changed, and users can tune the flags if they
want performance at the cost of supporting less hardware (e.g. to run code in
their own machine). This is not different to other architectures or to GCC,
where we default to the baseline that is officially supported by Debian.
Well, this is the same issue as SSE2 for i386. I would like to avoid
penalizing performances the whole population using
this arch just for a small percentage...
As example, two projects focusing on performance (Firefox and Chrome)
made that call.
Do we have a clue how many armhf systems don't have NEON?
So I think this should go back to RC and be fixed for stretch. BTW I think
Gianfranco had a patch for this, maybe he can attach it here.
The patch has not been tested. We are not sure we can force NEON with
the patch.
BTW on an unrelated note, llvm-toolchain-3.6 is out of testing!
Good! Thanks.
Sylvestre
