Re: [PATCH] x86/Intel: SHLD/SHRD have dual meaning

[Jan Beulich <jbeulich@suse.com>]

On 22.04.2024 09:43, Cui, Lili wrote:
>> On 22.04.2024 06:09, Hongtao Liu wrote:
>>> On Fri, Apr 19, 2024 at 5:29 PM Jan Beulich <jbeulich@suse.com> wrote:
>>>>
>>>> Since we uniformly permit D suffixes in Intel mode whenever in AT&T
>>>> mode an L suffix may be used, we need to be consistent with this.
>>> I think we need to forbid the D suffix for APX NDD SHL/SHR under Intel
>>> mode to avoid ambiguity.
>>
>> Hmm. Special casing just two insns is out of question imo (in fact that's what
>> is - unintentionally - partly happening prior to the change here).
>>
>>> Neither SHL (always SAL), nor SHR with a D suffix (Intel mode) is
>>> generated by GCC.
>>
>> While this may be deemed helpful, I actually view it as a mistake, even if only
>> from a cosmetic perspective: SHL is the main insn; SAL is merely an alias
>> (questionably using extension opcode 4 rather than 6 in gas).
>>
>> Still extending what you suggest - limiting the restriction to APX - may be an
>> option, albeit ...
>>
>>>> ---
>>>> The alternative, more intrusive and more risky (in terms of perceived
>>>> or even real regressions) route would be to mark the few insns which
>>>> permit suffixes even in Intel syntax, and reject suffix uses when
>>>> that indicator isn't set.
>>
>> ... as implied here I'd still consider this inconsistent.
>>
>> Furthermore the point in time when suffixes are processed off of the
>> incoming mnemonic is too early to know whether an insn is having an APX
>> representation. Hence doing said extension of what you suggest would likely
>> end up quite hacky.
>>
> 
> If the user messes them up, we can't figure out the original intention because the input is exactly the same, maybe we can find a suitable place to inform the user that shl + apx ndd does not support suffix?

See my other reply to Hongtao. Plus, as said before, I'd much prefer if we
wouldn't start special-casing any specific insns in this regard. (Which
isn't to say that may not end up being necessary here.)

Jan