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

[H.J. Lu]

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.

-- 
H.J.