RE: [PATCH v2 0/8] Support Intel APX EGPR

["Cui, Lili" <lili.cui@intel.com>]

> >> Subject: Re: [PATCH v2 0/8] Support Intel APX EGPR
> >>
> >> On 06.11.2023 15:20, Cui, Lili wrote:
> >>>> -----Original Message-----
> >>>> From: Jan Beulich <jbeulich@suse.com>
> >>>> Sent: Monday, November 6, 2023 3:30 PM
> >>>>
> >>>> On 03.11.2023 17:42, Cui, Lili wrote:
> >>>>> But if we want to merge bextr's vex and evex formats, we need to
> >>>>> support BMI&(BMI |( APX_F&x64))
> >>>>
> >>>> Maybe more like BMI&(<tbd>|APX_F), with further work (which I was
> >>>> considering
> >>>> anyway) towards x64 becoming a prereq to the increasing number of
> >>>> 64-bit- only features? (The <tbd> may well be BMI as you suggest,
> >>>> even if that reads a little odd.
> >>>>
> >>>
> >>> Yes, most VEX instructions don't require x64, but apx_f is x64
> >>> based. If the
> >> format "BMI&(BMI |( APX_F&x64))"  is complicated to implement or
> >> looks ugly, maybe we can handle x64 uniformly for apx_f in tc-i386.c.
> >>
> >> Well, some adjustment is needed there anyway, at the very least for
> >> the equivalent of e.g. the present handling of AVX|AVX512F or
> FMA|AVX512F.
> >> The goal wants to be to balance the amount of special casing code
> >> against complications in representing data in the opcode table. One
> question I have is:
> >> In how far is it necessary to actually represent APX_F in the BMI templates?
> >> There are two things triggering use of the EVEX encoding,
> >> iirc: Use of an extended register or NF. Use of an extended register
> >> is itself already dependent upon APX_F, and whatever the
> >> representation of NF is going to be, its parsing could be made dependent
> upon APX_F, too.
> >> No (strong) need then for the template to enforce APX_F yet another
> >> time, hopefully.
> >>
> >
> > In [patch 2/8] Support APX GPR32 with extend evex prefix,
> 
> Yet again patch 2/8, but this time the title reference at least clarifies you mean
> 3/8.
> 

I was all mixed up last night.

> > I only merged AMX's vex and evex formats (both vex and evex require x64),
> due to x64 reasons, BMI and other VEX instructions listed in 3.1.5 are not
> merged yet.
> 
> I see. Looking at patch 3 is next.
> 
> > NDD also triggers EVEX encoding.
> 
> But not for insns which were previously VEX-encoded?
> 
EVEX promote from VEX has no ND bit. You are right, there are only two factors extend register and NF.

> > There are some VEX instructions that support NF, their vex and evex cannot
> be merged.
> 
> Why not?
> 
Like bextr, new EVEX table has NF tag, I think we cannot merge them.

bextr, 0xf7, BMI, Modrm|CheckOperandSize|Vex128|Space0F38|VexVVVVSrc|SwapSources|No_bSuf|No_wSuf|No_sSuf, { Reg32|Reg64, Reg32|Reg64|Unspecified|BaseIndex, Reg32|Reg64 }
bextr, 0xf7, BMI|APX_F, Modrm|CheckOperandSize|EVex128|NF|Space0F38|VexVVVVSrc|SwapSources|No_bSuf|No_wSuf|No_sSuf, { Reg32|Reg64, Reg32|Reg64|Unspecified|BaseIndex, Reg32|Reg64 }

Lili.