[Qemu-devel] sparc: potentially incorrect if dynamic npc?

[Qemu-devel] sparc: potentially incorrect if dynamic npc?

[Artyom Tarasenko]

/* XXX: potentially incorrect if dynamic npc */
static void do_branch

There are few comments like this in target-sparc/translate.c .
In what case is it incorrect?

Can do_branch functions be called with dynamic npc?

[Qemu-devel] Re: sparc: potentially incorrect if dynamic npc?

[Blue Swirl]

On 4/20/10, Artyom Tarasenko <atar4qemu@googlemail.com> wrote:
> /* XXX: potentially incorrect if dynamic npc */
>  static void do_branch
>
>  There are few comments like this in target-sparc/translate.c .
>  In what case is it incorrect?
>
>  Can do_branch functions be called with dynamic npc?

Maybe, if there are two branches or jumps in a row. The last time I
checked, at least for successive branches we do the correct thing.
It's pretty obscure, no compiler will generate such sequences and
nobody with asm skills either.

Another corner case that IIRC we don't handle is where a jump
instruction is at the end of the page and its delay slot instruction
is at next page. I think one commit for Microblaze added support for
this case.

[Qemu-devel] Re: sparc: potentially incorrect if dynamic npc?

[Artyom Tarasenko]

2010/4/20 Blue Swirl <blauwirbel@gmail.com>:
> On 4/20/10, Artyom Tarasenko <atar4qemu@googlemail.com> wrote:
>> /* XXX: potentially incorrect if dynamic npc */
>>  static void do_branch
>>
>>  There are few comments like this in target-sparc/translate.c .
>>  In what case is it incorrect?
>>
>>  Can do_branch functions be called with dynamic npc?
>
> Maybe, if there are two branches or jumps in a row. The last time I
> checked, at least for successive branches we do the correct thing.
> It's pretty obscure, no compiler will generate such sequences and
> nobody with asm skills either.

Why not? That's how you can implement a very compact 'case'
operator in asm. It is used in OBP.

But you are right, this seems to work (at least with ss-5 and
ss-20 OBP).

> Another corner case that IIRC we don't handle is where a jump
> instruction is at the end of the page and its delay slot instruction
> is at next page.

Do you mean that npc is dynamic in this case, or that if
npc is dynamic and the instruction is at the end of the page we have
a problem?

> I think one commit for Microblaze added support for
> this case.

-- 
Regards,
Artyom Tarasenko

solaris/sparc under qemu blog: http://tyom.blogspot.com/

[Qemu-devel] Re: sparc: potentially incorrect if dynamic npc?

[Blue Swirl]

On 4/20/10, Artyom Tarasenko <atar4qemu@googlemail.com> wrote:
> 2010/4/20 Blue Swirl <blauwirbel@gmail.com>:
>
> > On 4/20/10, Artyom Tarasenko <atar4qemu@googlemail.com> wrote:
>  >> /* XXX: potentially incorrect if dynamic npc */
>  >>  static void do_branch
>  >>
>  >>  There are few comments like this in target-sparc/translate.c .
>  >>  In what case is it incorrect?
>  >>
>  >>  Can do_branch functions be called with dynamic npc?
>  >
>  > Maybe, if there are two branches or jumps in a row. The last time I
>  > checked, at least for successive branches we do the correct thing.
>  > It's pretty obscure, no compiler will generate such sequences and
>  > nobody with asm skills either.
>
>
> Why not? That's how you can implement a very compact 'case'
>  operator in asm. It is used in OBP.
>
>  But you are right, this seems to work (at least with ss-5 and
>  ss-20 OBP).
>
>
>  > Another corner case that IIRC we don't handle is where a jump
>  > instruction is at the end of the page and its delay slot instruction
>  > is at next page.
>
>
> Do you mean that npc is dynamic in this case, or that if
>  npc is dynamic and the instruction is at the end of the page we have
>  a problem?

The problem is with code generation in general (delay slot instruction
gets translated effectively before the branch), could be related to
dynamic npc too but the branch could be bn or ba as well.