[PATCH] x86: Don't check if AVX512 template requires AVX512VL

[Jan Beulich]

On 21.06.2023 18:06, H.J. Lu wrote:
> On Tue, Jun 20, 2023 at 11:52 PM Jan Beulich <jbeulich@suse.com> wrote:
>>
>> On 20.06.2023 18:44, H.J. Lu via Binutils wrote:
>>> --- a/gas/config/tc-i386.c
>>> +++ b/gas/config/tc-i386.c
>>> @@ -6288,11 +6288,10 @@ check_VecOperands (const insn_template *t)
>>>    /* Templates allowing for ZMMword as well as YMMword and/or XMMword for
>>>       any one operand are implicity requiring AVX512VL support if the actual
>>>       operand size is YMMword or XMMword.  Since this function runs after
>>> -     template matching, there's no need to check for YMMword/XMMword in
>>> -     the template.  */
>>> +     template matching, there's no need to check for YMMword/XMMword nor
>>> +     AVX512VL in the template.  */
>>>    cpu = cpu_flags_and (t->cpu_flags, avx512);
>>>    if (!cpu_flags_all_zero (&cpu)
>>> -      && !t->cpu_flags.bitfield.cpuavx512vl
>>>        && !cpu_arch_flags.bitfield.cpuavx512vl)
>>>      {
>>>        for (op = 0; op < t->operands; ++op)
>>
>> While the code change is correct afaict, I don't think we want it,
>> which is related to your imo wrong editing of the comment: For one we
>> check for the flag to be clear (unlike the xmm/ymm checks). And then
>> the check is there to skip a pointless (in that case) loop. IOW the
>> loop is only needed for templates which do not explicitly name
>> AVX512VL as a required feature; templates which do can't validly have
>> a mix of zmm and one or both of xmm/ymm in any one operand. In fact if
>> you grep i386-opc.tbl for "AVX612VL.*ZMM" you won't find any single
>> match.
> 
> That is why the check isn't needed.

I'm confused now: You explicitly want to make gas perform more work for
no gain? The flag being clear is the common case, and hence in the
common case we can avoid getting into this loop. So yes, as said, your
code change is functionally correct (as in: not breaking things), but
it is at the same time making overall behavior worse. Please revert.

Jan