[OT?] Unimplemented instructions on a G5?

[OT?] Unimplemented instructions on a G5?

[Pantelis Antoniou]

Hello

I've recently came across a very strange problem.

I have some PPC code with some inline assembly for implementing
saturated arithmetic.

The code runs fine on a 8xx target and on the G4 development host.

When I upgraded to a G5 running Gentoo the same code terminates
with an illegal instruction trap.

The offending instruction is the seemingly harmless:

  mcrxr crX

   (Move to condition register from XER).

Does anyone have any idea if IBM left this instruction
unimplemented. Searching the docs I've found nothing.

It's not a deal-breaker but it is inconvinient to not be
able to run the target binaries on the host.

If it is unimplemented any idea of how hard it will be to
trap and emulate it?

Regards

Pantelis


** Sent via the linuxppc-dev mail list. See http://lists.linuxppc.org/

Re: [OT?] Unimplemented instructions on a G5?

[Segher Boessenkool]

> The offending instruction is the seemingly harmless:
>
>  mcrxr crX
>
>   (Move to condition register from XER).
>
> Does anyone have any idea if IBM left this instruction
> unimplemented. Searching the docs I've found nothing.

Well, the newest PowerPC architecture docs say that it is
an optional insn, i.e., not all cpus might implement it.

I can't say whether the 970 implements it or not.

http://www-106.ibm.com/developerworks/eserver/pdfs/archpub1.pdf
Chapter 6.

> If it is unimplemented any idea of how hard it will be to
> trap and emulate it?

In Linux?

Last time I did this, it was easy.  Can't seem to find the
patch right now, though.

Emulating will be slow, of course (but I think you know that).


Segher


** Sent via the linuxppc-dev mail list. See http://lists.linuxppc.org/

Re: [OT?] Unimplemented instructions on a G5?

[David Edelsohn]

>> The offending instruction is the seemingly harmless:
>>
>> mcrxr crX
>>
>> (Move to condition register from XER).
>>
>> Does anyone have any idea if IBM left this instruction
>> unimplemented. Searching the docs I've found nothing.

>>>>> Segher Boessenkool writes:
Segher> Well, the newest PowerPC architecture docs say that it is
Segher> an optional insn, i.e., not all cpus might implement it.

Segher> I can't say whether the 970 implements it or not.

	mcrxr is obsolete and not implemented in PPC970.  Executing the
instruction will result in an illegal instruction trap.

David

** Sent via the linuxppc-dev mail list. See http://lists.linuxppc.org/

Re: [OT?] Unimplemented instructions on a G5?

[Pantelis Antoniou]

David Edelsohn wrote:

>>>The offending instruction is the seemingly harmless:
>>>
>>>mcrxr crX
>>>
>>>(Move to condition register from XER).
>>>
>>>Does anyone have any idea if IBM left this instruction
>>>unimplemented. Searching the docs I've found nothing.
>>>
>>>
>
>
>
>>>>>>Segher Boessenkool writes:
>>>>>>
>>>>>>
>Segher> Well, the newest PowerPC architecture docs say that it is
>Segher> an optional insn, i.e., not all cpus might implement it.
>
>Segher> I can't say whether the 970 implements it or not.
>
>	mcrxr is obsolete and not implemented in PPC970.  Executing the
>instruction will result in an illegal instruction trap.
>
>David
>
>
>
Well this much is obvious. Where can I find a list
of these obsolete instructions so that I can
implement them in the illegal instruction handler?

The IBM documentation for the 970 declares these instructions
as the previous architecture papers, i.e. as nothing special.

Regards

Pantelis


** Sent via the linuxppc-dev mail list. See http://lists.linuxppc.org/

Re: [OT?] Unimplemented instructions on a G5?

[Pantelis Antoniou]

Segher Boessenkool wrote:

>
>> The offending instruction is the seemingly harmless:
>>
>>  mcrxr crX
>>
>>   (Move to condition register from XER).
>>
>> Does anyone have any idea if IBM left this instruction
>> unimplemented. Searching the docs I've found nothing.
>
>
> Well, the newest PowerPC architecture docs say that it is
> an optional insn, i.e., not all cpus might implement it.
>
> I can't say whether the 970 implements it or not.
>
> http://www-106.ibm.com/developerworks/eserver/pdfs/archpub1.pdf
> Chapter 6.
>
Thanks this was the information I needed. Actually it's only
the mcrxr instruction that's unimplemented. Just my luck :-)

>> If it is unimplemented any idea of how hard it will be to
>> trap and emulate it?
>
>
> In Linux?
>
> Last time I did this, it was easy.  Can't seem to find the
> patch right now, though.
>
> Emulating will be slow, of course (but I think you know that).
>
>
> Segher
>
>
>
>
>
Regards

Pantelis


** Sent via the linuxppc-dev mail list. See http://lists.linuxppc.org/