how long to keep support for gcc on x86?

[Peter Wemm]

On Sun, Jan 13, 2013 at 2:53 PM, Nathan Whitehorn
<nwhitehorn at freebsd.org> wrote:
> On 01/13/13 14:48, John-Mark Gurney wrote:
>> Peter Wemm wrote this message on Sun, Jan 13, 2013 at 14:26 -0800:
>>> On Sun, Jan 13, 2013 at 12:29 PM, John-Mark Gurney <jmg at funkthat.com> wrote:
>>>> Adrian Chadd wrote this message on Sat, Jan 12, 2013 at 23:44 -0800:
>>>>>
>>>>> People are still ironing out kinks/differences with clang. Anyone
>>>>> saying otherwise is likely pushing an agenda. :-)
>>>>>
>>>>> Thus I think adding clang-only code to the system right now is very,
>>>>> very premature. There still seem to be reasons to run systems on GCC
>>>>> instead of clang.
>>>>>
>>>>> If you have a need for new instruction support, perhaps look at adding
>>>>> it to our base GCC for the time being?
>>>>
>>>> I did look at it briefly, but I don't know gcc's internals, and it would
>>>> take me 5+ hours to do it, while someone who does know gcc would take
>>>> abount a half an hour (just a guess)...  I don't have the free time I
>>>> used to, otherwise I would of done it by now..
>>>
>>> It seems to me that since clang is the default compiler for the
>>> platforms that have AES-NI that the following could be done:
>>>
>>> * get the inline AES-NI stuff in and debugged and solid.
>>> * .. without breaking the existing gcc-compatible code
>>> * once the support is solid, decide what the appropriate thing to do for gcc is.
>>>
>>> .. so long as the existing code doesn't get broken.
>>>
>>> Trying to do backwards compatibility port to gcc with a moving target
>>> has potential to be a work multiplier.
>>
>> I already have a gcc compatible version of an improved AES-NI for
>> amd64...  The real question is, do I improve things further by using
>> intrinsics which means we can share code between amd64 and i386 and get
>> great performance from both, or do I simply make a seperate version
>> for i386 that is gcc compatible, but not as good performance...
>>
>> Though a lot of this last little bit of performance questions isn't too
>> useful since the overhead of the crypto framework and geom introduces
>> a significant overhead already...
>>
>> I'm not too interesting in creating AES-NI v2 module and having two
>> versions that do the same thing just because of a compiler issue...
>>
>> So I'm going to go with the plan of making an i386 and gcc compatible
>> version... it'll still be a 4x+ performance over the existing code...
>> This also means we could back port it to 9-stable if we wanted to...
>>
>> Thanks for the input...
>>
>
> This also raises the interesting question of whether we want to bother
> supporting things like AES-NI on i386 at all. It's a legacy/embedded
> architecture at this point, in my opinion, and the people who run it
> probably don't care about fancy new features like this. A related
> question is whether we want to have any clang-only features in the kernel...

It wasn't so much an issue of being clang-only, but rather the antique
gcc+binutils not having full support on i386.  The code would work
fine with later versions of gcc/binutils, intel's compiler and clang.
Just not with the old gcc + old binutils on i386.


-- 
Peter Wemm - peter at wemm.org; peter at FreeBSD.org; peter at yahoo-inc.com; KI6FJV
bitcoin:188ZjyYLFJiEheQZw4UtU27e2FMLmuRBUE