On Tue, Dec 01, 2015 at 06:27:17PM +0100, Paolo Bonzini wrote:
> On 01/12/2015 18:20, Richard Henderson wrote:
> >>
> >> X86XSaveArea will be used only when loading/saving state using
> >> xsave, not for executing regular instructions.
> >
> > ... like the regular instruction xsave?
> >
> > https://patchwork.ozlabs.org/patch/493318/
>
> Right, but that's a helper anyway.
>
> >> In X86CPU, the
> >> data is already stored as XMMReg unions (the one with the
> >> XMM_[BWDQ] helpers).
> >
> > Of course. But those unions are arranged to be in big-endian format on
> > big-endian hosts. So we need to swap the data back to little-endian
> > format for storage into guest memory.
>
> Yes, you can use byte moves with XMM_B (more obvious), or stq_le_p with
> XMM_Q (faster I guess---though the compiler might optimize the former on
> little-endian hosts). Either works with an uint8_t[] destination.
stq_le_p() (more exactly, stq_p()) is exactly what is already
done by kvm_{get,put}_xsave(), using uint8_t pointers.
BTW, if we are going to implement xsave in TCG, the
X86CPU<->xsave translation logic in kvm_{get,put}_xsave() could
be moved to generic code and reused by TCG instead of being
reimplemented.
--
Eduardo
