[Qemu-devel] Re: [Qemu-commits] [COMMIT 968a40f] microblaze: Trap on unaligned data accesses.

[Qemu-devel] Re: [Qemu-commits] [COMMIT 968a40f] microblaze: Trap on unaligned data accesses.

[Blue Swirl]

On Thu, Sep 3, 2009 at 2:11 PM, Anthony Liguori<aliguori@us.ibm.com> wrote:
> From: Edgar E. Iglesias <edgar.iglesias@gmail.com>

> +void helper_memalign(uint32_t addr, uint32_t dr, uint32_t wr, uint32_t size)
> +{
> +    uint32_t mask;
> +
> +    switch (size) {
> +        case 4: mask = 3; break;
> +        case 2: mask = 1; break;
> +        default:
> +        case 1: mask = 0; break;
> +    }

For additional performance (and to avoid the check for mask == 0 case
;-)), you could perform most of the math above during translation.

Sparc32 actually avoids most of these explicit tests, but I still
can't understand Fabrice's magic that made it possible.

[Qemu-devel] Re: [Qemu-commits] [COMMIT 968a40f] microblaze: Trap on unaligned data accesses.

[Edgar E. Iglesias]

On Thu, Sep 03, 2009 at 09:10:40PM +0300, Blue Swirl wrote:
> On Thu, Sep 3, 2009 at 2:11 PM, Anthony Liguori<aliguori@us.ibm.com> wrote:
> > From: Edgar E. Iglesias <edgar.iglesias@gmail.com>
> 
> > +void helper_memalign(uint32_t addr, uint32_t dr, uint32_t wr, uint32_t size)
> > +{
> > +    uint32_t mask;
> > +
> > +    switch (size) {
> > +        case 4: mask = 3; break;
> > +        case 2: mask = 1; break;
> > +        default:
> > +        case 1: mask = 0; break;
> > +    }
> 
> For additional performance (and to avoid the check for mask == 0 case
> ;-)), you could perform most of the math above during translation.

Thanks, I beleive part of what you say is already done. The case 1
is there for completeness but translate.c doesn't call the helper for
size 1.

I'll move the computing of the mask into translate.c.

Thanks,
Edgar

> 
> Sparc32 actually avoids most of these explicit tests, but I still
> can't understand Fabrice's magic that made it possible.