Re: [PATCH V2] Support APX NF

[Jan Beulich <jbeulich@suse.com>]

On 12.03.2024 07:12, Cui, Lili wrote:
> 
> 
>> -----Original Message-----
>> From: Jan Beulich <jbeulich@suse.com>
>> Sent: Monday, March 11, 2024 10:10 PM
>> To: Cui, Lili <lili.cui@intel.com>
>> Cc: hjl.tools@gmail.com; binutils@sourceware.org
>> Subject: Re: [PATCH V2] Support APX NF
>>
>> On 11.03.2024 14:54, Cui, Lili wrote:
>>>> -----Original Message-----
>>>> From: Jan Beulich <jbeulich@suse.com>
>>>> Sent: Friday, March 8, 2024 5:36 PM
>>>>
>>>> On 04.03.2024 09:15, Cui, Lili wrote:
>>>>> @@ -7218,6 +7229,11 @@ parse_insn (const char *line, char
>> *mnemonic,
>>>> bool prefix_only)
>>>>>  		  /* {rex2} */
>>>>>  		  i.rex2_encoding = true;
>>>>>  		  break;
>>>>> +		case Prefix_NF:
>>>>> +		  /* {nf} */
>>>>> +		  i.has_nf = true;
>>>>> +		  i.encoding = encoding_evex;
>>>>> +		  break;
>>>>
>>>> It's not quite as easy, I'm afraid: Have you thought of the "{vex} {nf} ..."
>>>> case? (I think I previously indicated that their combination,
>>>> actually in either order, needs properly rejecting.) Without having
>>>> spent much thought on it, perhaps it would suffice to check here that
>>>> the field is still encoding_default, and leave the value alone
>>>> otherwise (in order to reject bad combinations elsewhere).
>>>>
>>> Oh, good point. Do you think it's ok to report bad outside of "switch"?
>>>
>>>                 case Prefix_NF:
>>>                   /* {nf} */
>>>                   i.has_nf = true;
>>> -                 i.encoding = encoding_evex;
>>> +                 if (i.encoding == encoding_default)
>>> +                   i.encoding = encoding_evex;
>>>                   break;
>>>                 case Prefix_NoOptimize:
>>>                   /* {nooptimize} */
>>>                   i.no_optimize = true;
>>>                   break;
>>>                 default:
>>>                   abort ();
>>>                 }
>>> +             if (i.has_nf && i.encoding != encoding_evex)
>>> +               {
>>> +                 as_bad (_("{nf} cannot be combined with {vex}/{vex3}"));
>>> +                 return NULL;
>>> +               }
>>
>> Where the checking is done is secondary to me. My primary point is that the
>> above isn't correct: There are more encoding variants than just vex, vex3, and
>> evex.
>>
>> That said, reporting the error right here may lead to multiple identical errors
>> being reported for a single line of input. This wants avoiding, at which point
>> placing the check here is ruled out.
>>
> 
> Sorry,  I didn’t get you here, do you think the current implementation is wrong? I think for parse_insn stage i.encodin only has those encoding variants,

Hmm, you're right - operands weren't parsed yet at this point. While ...

> do you think we need to move further back? Maybe put them together with REX/REX2 prefix checking?
> 
>       /* Check for explicit REX prefix.  */
>       if (i.prefix[REX_PREFIX] || i.rex_encoding)
>         {
>           as_bad (_("REX prefix invalid with `%s'"), insn_name (&i.tm));
>           return;
>         }
> 
>       /* Check for explicit REX2 prefix.  */
>       if (i.rex2_encoding)
>         {
>           as_bad (_("{rex2} prefix invalid with `%s'"), insn_name (&i.tm));
>           return;
>         }

... this would overall seem to be a more logical place to have such checks,
I'd like to take back my earlier comment then: By having the check in
parse_insn(), .insn handling is then also covered right away.

Jan