Re: [PATCH 8/8] KVM: x86: simplify kvm_apic_map

["Radim Krčmář" <rkrcmar@redhat.com>]

2015-01-30 10:18+0100, Paolo Bonzini:
> On 29/01/2015 22:48, Radim Krčmář wrote:
> > +static inline bool
> > +apic_logical_id(struct kvm_apic_map *map, u32 ldr, u16 *cid, u16 *lid)
> >  {
> > +	switch (map->mode) {
> > +	case KVM_APIC_MODE_XAPIC_FLAT:
> > +		*cid = 0;
> > +		*lid = ldr & 0xff;
> > +		return true;
> > +	case KVM_APIC_MODE_XAPIC_CLUSTER:
> > +		*cid = (ldr >> 4) & 0xf;
> > +		*lid = ldr & 0xf;
> > +		return true;
> > +	case KVM_APIC_MODE_X2APIC:
> > +		*cid = ldr >> 16;
> > +		*lid = ldr & 0xffff;
> > +		return true;
> > +	}
> 
> We need some optimization here.  You can make the defines equal to the
> size of the lid: CLUSTER = 1 << 3, FLAT = 1 << 2, X2APIC = 1 << 4:
> 
> 	BUILD_BUG_ON(...FLAT != 4);
> 	BUILD_BUG_ON(...CLUSTER != 8);

(Swapped.)

> 	BUILD_BUG_ON(...X2APIC != 16);

(Check the mode and return false here.)

> 	lid_bits = mode;
> 	cid_bits = mode & (16 | 4);
> 	lid_mask = (1 << lid_bits) - 1;
> 	cid_mask = (1 << cid_bits) - 1;
> 
> 	*cid = (ldr >> lid_bits) & cid_mask;
> 	*lid = ldr & lid_mask;

Would jump predictor fail on the switch?  Or is size of the code that
important?  This code is shorter, but is going to execute far more
operations, so I think it would be slower ... (And harder to read.)

> Please move this to lapic.c since you are at it.

Ok.