Re: RFC: Petition Intel/AMD to add POPF_IF insn

[Denys Vlasenko <dvlasenk@redhat.com>]

On 08/17/2016 09:13 PM, Andy Lutomirski wrote:
> On Wed, Aug 17, 2016 at 12:01 PM, Linus Torvalds
> <torvalds@linux-foundation.org> wrote:
>> On Aug 17, 2016 11:41 AM, "Denys Vlasenko" <dvlasenk@redhat.com> wrote:
>>>
>>> OTOH 5 years will inevitably pass.
>>
>> Yes. But in five years, maybe we'll have a popf that is faster anyway.
>>
>> I'd actually prefer that in the end. The problem with popf right now seems
>> to be mainly that it's effectively serializing and does stupid things in
>> microcode. It doesn't have to be that way. It could actually do much better,
>> but it hasn't been a high enough priority for Intel.
>>
>
> It wouldn't surprise me if that were easier said than done.  popf
> potentially changes AC, and AC affects address translation.  popf also
> potentially changes IOPL, and I don't know whether Intel chips track
> IOPL in a way that lets them find all the dependent instructions
> without serializing.  But maybe their pipeline is fancy enough.

Exactly. And more:

POPF potentially changes TF (and it works even in CPL3).
POPD changes DF - must serialize versus string insns.
POPF changes NT - must serialize versus IRET insns.
POPF changes VIF, from a different bit in a popped value,
and under a rather complex conditions.

Intel's documentation has a pseudo-code for the instructions.
For POPF, that pseudo-code is two pages long...