[PATCH] x86/APX: Don't promote AVX512 instructions out of APX spec

[Jan Beulich]

On 23.09.2024 13:14, Cui, Lili wrote:
> 
> 
>> -----Original Message-----
>> From: Jan Beulich <jbeulich@suse.com>
>> Sent: Monday, September 23, 2024 4:30 PM
>> To: Cui, Lili <lili.cui@intel.com>
>> Cc: Binutils <binutils@sourceware.org>; H.J. Lu <hjl.tools@gmail.com>
>> Subject: Re: [PATCH] x86/APX: Don't promote AVX512 instructions out of APX
>> spec
>>
>> On 23.09.2024 09:51, Cui, Lili wrote:
>>> Jan, I'm sorry that I didn't notice you checking in these patches at first,
>> although I have been against this kind of promotion in the previous discussion.
>> But it's really a bit radical to make this kind of out-of-spec promotion by default.
>> Would you rather put them under options? After all, these conversions are very
>> clever.
>>
>> I continue to think they want doing by default, with possibly a way to turn them
>> off (compiler mode) and/or diagnose them (perhaps no more than "info"
>> level).
>>
> 
> The benefit of this optimization is that when the hand-written assembler makes a mistake in writing few VEX instructions, we can promote it to the EVEX instructions. The cost is to let the compiler add a separate option when compiling or define different behaviors for the compiler and the hand-written assembler mode (If there is a way to do this, inline assembler and hand-written assembly behavior is inconsistent, this inconsistent behavior will confuse people). Other compilers also strictly enforce errors for this illegal Egpr usage. We cannot promote all illegal Egpr usage, if there is such an improvement,

I'm relatively sure this could be arranged for in many of the cases, just not
with single-insn replacements. I didn't want to go _that_ far ...

Jan 

> it may be worth making such a change.
>  
> Thanks,
> Lili.
> 
>> Jan