[Pixman] [PATCH 00/10] Cleanups to CPU detection
Alan Coopersmith
alan.coopersmith at oracle.com
Fri Jun 29 14:40:02 PDT 2012
On 06/29/12 02:36 PM, sandmann at cs.au.dk wrote:
> Alan Coopersmith <alan.coopersmith at oracle.com> writes:
>
>> On 06/29/12 01:44 PM, Søren Sandmann Pedersen wrote:
>>> I was looking at making use of some of the newer x86 SIMD instruction
>>> sets and realized that (a) we don't ever call cpuid on x86-64, we just
>>> assume that MMX and SSE2 are present,
>>
>> I thought the amd64 ABI guaranteed MMX & SSE2 would always be present - is
>> that not the case?
>
> The problem is not so much that assumption, but the fact that for newer
> instruction sets such as SSSE3 and AVX, we do need to call cpuid even on
> x86-64. So the main point is just preparing the code for that.
>
> But in fact, I'm pretty sure I saw in some Knight's Corner (which is
> x86-64) manual, linked from here:
>
> http://software.intel.com/en-us/forums/showthread.php?t=105443
>
> that it doesn't support SSE and MMX, only the new 512 bit vector
> instructions. I can't find it now though, so maybe I was
> hallucinating. They do say that that chip is not binary compatible with
> other x86 chips, and in the instruction manual chapter 4.8, they say
> that the state of mm0 through mm7 and xmm0 through xmm7 is "NA".
Ah, okay. It should't hurt, and if it may be useful going forward,
then go for it.
--
-Alan Coopersmith- alan.coopersmith at oracle.com
Oracle Solaris Engineering - http://blogs.oracle.com/alanc
More information about the Pixman
mailing list