Proposal: Let's drop i386
On Sun, May 13, 2018 at 9:34 PM, Matthias Klose <doko at ubuntu.com> wrote:
> On 13.05.2018 05:00, Henri Sivonen wrote:
>> On Thu, May 10, 2018 at 11:25 PM, Thomas Ward <teward at thomas-ward.net> wrote:
>>> However, killing i386 support globally could introduce issues, including
>>> but not limited to certain upstream softwares having to go away
>>> entirely, due to the interdependency or issues with how certain apps
>>> work (read; Wine, 32-bit support, 64-bit support being flaky, and
>>> Windows apps being general pains in that they work on 32bit but not
>>> always on 64-bit).
>> If 32-bit x86 support becomes mainly a thing that's run on x86_64
>> hardware as a compatibility measure for things like Wine, it would
>> make sense to bring the instruction set baseline to the x86_64 level.
>> Specifically, it would make sense to compile the 32-bit x86 packages
>> with SSE2 unconditionally enabled.
>> This would mean dropping support for Pentium Pro and earlier or Athlon
>> XP and earlier, but it's pretty sad to leave all that performance on
>> the table in order to support the few computers still in use that have
>> Pentium Pro or earlier or Athlon XP or earlier.
>> As upstream software assumes SSE2 as the baseline, it will be less and
>> less a run-time check and compiling software without SSE2 will mean
>> shipping it in a damaged form performance-wise.
> I disagree, until you provide data how many packages fail to build, at least in
> the testsuites, when run without the extra x87 precision bits.
I don't have this data, but considering that SSE2 is a mandatory part
of x86_64, it seems implausible that packages would be
SSE2-intolerant. Considering that x86_64 defaults to SSE2
floating-point math (or does Ubuntu override this?) and considering
that ARM doesn't have x87 available, it seems implausible that
packages would rely on x87. (On the contrary, since e.g. Firefox and
Chromium upstreams don't do non-SSE2 x86 builds anymore, it seems more
plausible that there exist packages whose upstream doesn't test x87
floating-point math anymore.)
hsivonen at hsivonen.fi