Jens Nerche wrote:
>
> >That paper is the closest thing I have to a summary.
> You recommend for branches, especial for those with calculated targets,
> the single-step-technique. This means: set flag, return to guest, jump into
> monitor again and resume. Lots of switches. And switches (ring0/3) are
> expensive on x86, also because we save/restore the whole CPU state.
> Is it cheaper to emulate all branches in monitor?
> Yes, in my page table virtualization code I did also single stepping, but
> because I didn't have an emulator in guest side. Now we have one, perhaps
> someone can do some measurements?
Given current thinking, best to just emulate the branch instruction
in the monitor, since we're already there after receiving
an interrupt from the virtualized instruction. I think at
the time I was thinking in terms of being lazy and not
emulating any more than we have to. But since we have
to be able to emulate instructions anyways, we should
do it in the monitor.
Good call.
-Kevin
