assorted PPC questions

[Jan Beulich]

On 17.08.2020 15:28, Alan Modra wrote:
> On Mon, Aug 17, 2020 at 10:19:28AM +0200, Jan Beulich wrote:
>> On 11.08.2020 11:02, Alan Modra wrote:
>>> On Wed, Aug 05, 2020 at 09:38:35AM +0200, Jan Beulich wrote:
>>>> Wouldn't availability of pseudos implying %cr0 as first (i.e.
>>>> destination) operand better be consistent throughout the opcode set?
>>>
>>> Are you saying every instruction ought to have a record bit?  ie. that
>>> there ought to be xxx. for every xxx?
>>
>> No, the question wasn't about the . suffix. Instead what I'm wondering
>> is why CMP*, VCMP*, and XVTLSBB have pseudos with the first operand
>> omitted (implying %cr0), while e.g. FCMP* doesn't (and there are quite
>> a few more).
> 
> I'm not sure.  Possibly because you might expect that with a fcmpu
> form that allowed the BF field to be omitted, the implied CR field
> should be cr1 in line with with the record form of other fp insns,
> eg. fmr. sets cr1.  That conflicts with the general rule that an
> omitted field has a zero value.  (Yes, we now have a few exceptions,
> see *invalid < 0 tests in ppc-opc.c)

Well, then maybe FCMP* was a bad example. What about CMPRB? Or possibly
some vector-scalar insns like xscmp*?

Jan