Re: [PATCH] Correct cpu_flags for CPU_ANY_AVX512BW_FLAGS

["H.J. Lu" <hjl.tools@gmail.com>]

On Thu, Jun 23, 2022 at 9:31 AM H.J. Lu <hjl.tools@gmail.com> wrote:
>
> On Thu, Jun 23, 2022 at 9:08 AM Jan Beulich <jbeulich@suse.com> wrote:
> >
> > On 23.06.2022 18:04, H.J. Lu wrote:
> > > On Thu, Jun 23, 2022 at 1:08 AM Cui, Lili <lili.cui@intel.com> wrote:
> > >>
> > >>
> > >>
> > >>> -----Original Message-----
> > >>> From: Jan Beulich <jbeulich@suse.com>
> > >>> Sent: Thursday, June 23, 2022 3:30 PM
> > >>> To: Cui, Lili <lili.cui@intel.com>
> > >>> Cc: binutils@sourceware.org; hjl.tools@gmail.com
> > >>> Subject: Re: [PATCH] Correct cpu_flags for CPU_ANY_AVX512BW_FLAGS
> > >>>
> > >>> On 23.06.2022 09:14, Cui,Lili wrote:
> > >>>> This patch is to correct cpu_flags for CPU_ANY_AVX512BW_FLAGS.
> > >>>> Make check-gas is ok.
> > >>>> Ok for master?
> > >>>>
> > >>>> Thanks,
> > >>>> Lili.
> > >>>>
> > >>>> Change it from "CPU_ANY_AVX512_FP16_FLAGS" to "CpuAVX512_FP16"
> > >>>
> > >>> You don't say what's wrong with what there is right now. I for one think what
> > >>> we have is correct: If one disables BW, _all_ its dependents should be
> > >>> disabled as well, which transitively means all dependents of FP16.
> > >>>
> > >> Hi Jan,
> > >> Agree with that " If one disables BW, _all_ its dependents should be disabled as well ".
> > >> I saw the form of "CPU_ANY_AVX512F_FLAGS" today, and to be consistent with it, I thought "CPU_ANY_AVX512_FP16_FLAGS" was wrong at that time. But after reading your email, I found "CPU_ANY_AVX512F_FLAGS" to be the only one that uses cpu_flags instead of CPU_ANY* .
> > >>
> > >> { "CPU_ANY_AVX512F_FLAGS",
> > >> "CpuAVX512F|CpuAVX512CD|CpuAVX512ER|CpuAVX512PF|CpuAVX512DQ|CPU_ANY_AVX512BW_FLAGS|CpuAVX512VL|CpuAVX512IFMA|CpuAVX512VBMI|CpuAVX512_4FMAPS|CpuAVX512_4VNNIW|CpuAVX512_VPOPCNTDQ|CpuAVX512_VBMI2|CpuAVX512_VNNI|CpuAVX512_BITALG|CpuAVX512_BF16|CpuAVX512_VP2INTERSECT" },
> > >>
> > >> Regards,
> > >> Lili.
> > >>
> > >>> Jan
> > >>>
> > >>>> opcodes/
> > >>>>
> > >>>>     * i386-gen.c (cpu_flag_init): Change cpu_flags of
> > >>>>     CPU_ANY_AVX512BW_FLAGS.
> > >>>> ---
> > >>>>  opcodes/i386-gen.c | 2 +-
> > >>>>  1 file changed, 1 insertion(+), 1 deletion(-)
> > >>>>
> > >>>> diff --git a/opcodes/i386-gen.c b/opcodes/i386-gen.c index
> > >>>> d18a7d2754..5dc224efdb 100644
> > >>>> --- a/opcodes/i386-gen.c
> > >>>> +++ b/opcodes/i386-gen.c
> > >>>> @@ -382,7 +382,7 @@ static initializer cpu_flag_init[] =
> > >>>>    { "CPU_ANY_AVX512DQ_FLAGS",
> > >>>>      "CpuAVX512DQ" },
> > >>>>    { "CPU_ANY_AVX512BW_FLAGS",
> > >>>> -    "CpuAVX512BW|CPU_ANY_AVX512_FP16_FLAGS" },
> > >>>> +    "CpuAVX512BW|CpuAVX512_FP16" },
> > >>>>    { "CPU_ANY_AVX512VL_FLAGS",
> > >>>>      "CpuAVX512VL" },
> > >>>>    { "CPU_ANY_AVX512IFMA_FLAGS",
> > >>
> > >
> > > Since CPU_ANY_AVX512_FP16_FLAGS is the same as CpuAVX512_FP16,
> > > this change is nop.  On the other hand, CPU_ANY shouldn't be used to define
> > > another CPU_ANY.
> >
> > Why? Imo that's exactly how it should be, as long as the purpose of
> > CPU_ANY_* is the use for processing of negative .arch directives.
> >
>
> Assuming AVX512_FP16 also requires BMI2, when AVX512F is disabled,
> AVX512_FP16 should be disabled, but not BMI2.
>

I was wrong.  The current code isn't wrong since

 { "CPU_ANY_AVX512BW_FLAGS",
   "CpuAVX512BW|CPU_ANY_AVX512_FP16_FLAGS" },

indicates that disabling AVX512BW will also disable AVX512BW +
AVX512F16 + any ISAs which
depend on AVX512FP16.   If there is another ISA which depends on
AVX512FP16, we don't
need to change CPU_ANY_AVX512BW_FLAGS.

-- 
H.J.