Re: [Qemu-devel] [PATCH 3/3] target-i386: replace cpuid_*features fields with a feature word array

[Eduardo Habkost]

On Fri, Dec 14, 2012 at 06:47:20PM +0100, Igor Mammedov wrote:
> On Fri, 14 Dec 2012 15:36:22 -0200
> Eduardo Habkost <address@hidden> wrote:
> 
> > On Fri, Dec 14, 2012 at 06:20:41PM +0100, Andreas Färber wrote:
> > > Am 14.12.2012 17:52, schrieb Eduardo Habkost:
> > > > On Fri, Dec 14, 2012 at 04:14:32PM +0100, Andreas Färber wrote:
> > > >> Am 12.12.2012 23:22, schrieb Eduardo Habkost:
> [...]
> > > 
> > > >> The clock code using first_cpu looks solvable; what about CR4 and MSR
> > > >> helpers, how performance-sensitive are they? (if they're not yet using
> > > >> X86CPU for something else)
> > > > 
> > > > I guess any CPU-state code inside QEMU is not performance-sensitive, as
> > > > it woud already require switching between KVM kernelspace and QEMU
> > > > userspace.
> > > 
> > > I mean target-i386/[misc_]helper.c and thus TCG, IIUC. :)
> > 
> > Oh, right. I wonder how much performance impact it would have, if people
> > are already using TCG.
> > 
> > Anyway, would this really have any impact at all? I mean:
> > ENV_GET_CPU(env) is basically subtracing an constant offset from 'env'.
> > So I expect similar code to be generated, just using a different offset
> > from 'env' to get the cpuid_features field.
> ENV_GET_CPU(env) does dynamic_cast which is expensive.

Oh, I didn't notice that. So the alternatives I see are:
- Use ENV_GET_CPU() and risk performance problems;
- Write a FAST_ENV_GET_CPU() macro for performance-sensitive code, that
  doesn't use dynamic_cast;
- Keep the fields on CPUX86State and move them only after we change the
  TCG code to use the QOM CPU object.

Personally, I prefer the third option. Moving fields from CPUArchState
before making most code use the CPU QOM objects instead of CPUArchState
sounds like a painful task.

-- 
Eduardo