Re: [Fwd: Re: [Qemu-devel] [PATCH] SVM support]

[Alexander Graf]

On Sep 19, 2007, at 4:39 PM, Jocelyn Mayer wrote:

> On Wed, 2007-09-19 at 12:56 +0200, Alexander Graf wrote:
> > On Sep 19, 2007, at 1:28 AM, Paul Brook wrote:
> >
> > > > > > Ok, I will try to shift the intercepts in an uint_64 flags  
> > > > > > variable
> > > > > > in the TB.
> > > >
> > > > OK, great. Having 64 bits may also help for additional (ie  
> > > > future...)
> > > > features in PowerPC 64 emulation.
> > >
> > > Maybe worth letting the target say whether it needs 32 or 64-bit
> > > flags.
> > > The flag lookup is likely to be on a hot path.
> > >
> > > Paul
> >
> > here comes the patch.
>
> The patch looks correct to me, I'm just wondering why a cast to 32  
> bits
> is needed in mips code ?


It's not needed - I just did it so one can see that there is  
potential that is not used right now and someone should think about  
it. I'm not too much into mips so I have no clues if they have any  
use for more that 32 bits of flags.

> I may be wrong, but it seems to me that TB flags are only used to be
> compared in tb_find_fast and tb_find_slow, which are not a very fast
> path (we come here when we cannot jump directly to the next tb and are
> not in the most time critical code) then we can afford adding a few  
> asm
> instructions (I would say max 2 ?) to replace a 32 bits comparison  
> with
> a 64 bits one. My feeling is that the performance loss here won't be
> sensible, but that may need to be proved.

I actually don't even believe the difference can be measured at all,  
which is why I seperated the patches.

Cheers,

Alex