On Mon, Jan 19, 2015 at 9:41 PM, Reindl Harald <h.reindl(a)thelounge.net> wrote:
Am 19.01.2015 um 21:34 schrieb drago01:
>
> On Mon, Jan 19, 2015 at 3:42 PM, Reindl Harald <h.reindl(a)thelounge.net>
> wrote:
>>
>>
>> Am 19.01.2015 um 15:28 schrieb Gerd Hoffmann:
>>>>>
>>>>>
>>>>> Would that make it impossible to run fedora on sse-only i686 CPUs?
>>>>
>>>>
>>>>
>>>> Yes, but test coverage for those CPUs is already rather poor, so I
>>>> don't
>>>> expect them to work anymore.
>>>
>>>
>>>
>>> See other replies, people are running such machines.
>>>
>>> Given x86_64 exists for many years and even embedded (see aarch64) is
>>> moving to 64bit these days i386 clearly is for old legacy hardware. If
>>> you need performance you don't run a i386 machine.
>>>
>>> So, why bother raising the bar for i386 by requiring more cpu features?
>>> It doesn't help anyone.
>>
>>
>>
>> "rpmrc" has optflags: for different architectures
>>
>> so why not optimize only "optflags: x86_64" for more recent hardware?
>> since i686 is a complete different build anyways it won't matter
>>
>> on my private builders it's "-mmmx -msse2 -msse3 -msse4.1 -msse4.2
-maes
>> -mfpmath=sse"
>
>
> Because that doesn't make sense for multiple reasons:
you show none below because it said "keep i686 flags unchanged"
> 1) mmx is basically obsolete
in theory
And in practice. Why would you want to use mmx if you have sse2+
available? (you don't).
> 2) sse2 is the default fp ABI on x86_64 anyway and thus supported
by
> *all* x86_64 cpus
so be it
> 3) Requiring sse4.2 or even 4.1 would exclude a lot of hardware
did i propose that?
It wasn't clear from your mail what your proposal actually is.
i said "private builders" to show that keep supporting >
10 years old
hardware is just silly because you waste HW capabilities - noweher did i
propose *that* falgs
> 4) You don't just throw in some compiler flags and assume it will gain
> you anything (you'd probably have to either turn on O3 or
> vectorization as well)
-O3 is the default for my builds starting with 2006
> .. most applications where it really matters do
> detection at runtime (even glibc does that)
the topic is still: hwta can be changed in the x86_64 flags *without* break
old 32bit crap some people rely for several reasons
OK .. well there is no reason why the flags have to be shared between
x86_64 and the "32bit crap" ... if you stop carring about 32bit
*hardware* we could use sse2 just fine for i686 (people using 32bit
applications on top of x86_64 would benefit from it).