[PATCH 01/10] Support Intel AVX-IFMA
H.J. Lu
hjl.tools@gmail.com
Mon Oct 24 19:09:51 GMT 2022
On Sun, Oct 23, 2022 at 10:53 PM Jiang, Haochen <haochen.jiang@intel.com> wrote:
>
> > -----Original Message-----
> > From: Jan Beulich <jbeulich@suse.com>
> > Sent: Friday, October 14, 2022 5:53 PM
> > To: Jiang, Haochen <haochen.jiang@intel.com>
> > Cc: hjl.tools@gmail.com; Wang, Hongyu <hongyu.wang@intel.com>;
> > binutils@sourceware.org
> > Subject: Re: [PATCH 01/10] Support Intel AVX-IFMA
> >
> > On 14.10.2022 11:12, Haochen Jiang wrote:
> > > From: wwwhhhyyy <hongyu.wang@intel.com>
> > >
> > > x86: Support Intel AVX-IFMA
> > >
> > > Intel AVX IFMA instructions are marked with CpuVEX_PREFIX, which is
> > > cleared by default. Without {vex} pseudo prefix, Intel IFMA instructions
> > > are encoded with EVEX prefix. {vex} pseudo prefix will turn on VEX
> > > encoding for Intel IFMA instructions.
> >
> > I firmly object to the proliferation of this mis-feature. As expressed
> > before for AVX-VNNI, as long as the user has disabled AVX512 (or
> > respective sub-features thereof), there should be no need to use {vex} in
> > the source code. There's also no reason at all to make the disassembler
> > print {vex} prefixes - we don't do so for any other insns (apart from
> > AVX-VNNI) where an ambiguity exists between their VEX and EVEX encodings
> > (when none of the EVEX-specific features is used).
> >
> > I actually have a patch queued to undo the odd behavior for AVX-VNNI, at
> > least on the assembler side (which also drops the PseudoVexPrefix
> > attribute).
>
> Has rebased the patch to latest trunk and removed PseudoVexPrefix in table.
> Also added some testcases just like how your patch did.
>
> >
> > > --- a/opcodes/i386-dis.c
> > > +++ b/opcodes/i386-dis.c
> > > @@ -1526,6 +1526,8 @@ enum
> > > VEX_W_0F385E_X86_64_P_3,
> > > VEX_W_0F3878,
> > > VEX_W_0F3879,
> > > + VEX_W_0F38B4,
> > > + VEX_W_0F38B5,
> > > VEX_W_0F38CF,
> > > VEX_W_0F3A00_L_1,
> > > VEX_W_0F3A01_L_1,
> > > @@ -6293,8 +6295,8 @@ static const struct dis386 vex_table[][256] = {
> > > { Bad_Opcode },
> > > { Bad_Opcode },
> > > { Bad_Opcode },
> > > - { Bad_Opcode },
> > > - { Bad_Opcode },
> > > + { VEX_W_TABLE (VEX_W_0F38B4) },
> > > + { VEX_W_TABLE (VEX_W_0F38B5) },
> > > { "vfmaddsub231p%XW", { XM, Vex, EXx }, PREFIX_DATA },
> > > { "vfmsubadd231p%XW", { XM, Vex, EXx }, PREFIX_DATA },
> > > /* b8 */
> > > @@ -7599,6 +7601,16 @@ static const struct dis386 vex_w_table[][2] = {
> > > /* VEX_W_0F3879 */
> > > { "vpbroadcastw", { XM, EXw }, PREFIX_DATA },
> > > },
> > > + {
> > > + /* VEX_W_0F38B4 */
> > > + { Bad_Opcode },
> > > + { "%XV vpmadd52luq", { XM, Vex, EXx }, PREFIX_DATA },
> > > + },
> > > + {
> > > + /* VEX_W_0F38B5 */
> > > + { Bad_Opcode },
> > > + { "%XV vpmadd52huq", { XM, Vex, EXx }, PREFIX_DATA },
> > > + },
> >
> > Irrespective of the aspect mentioned at the top I think this is yet
> > another case where VEX and EVEX table entries can be shared. This would
> > (if the {vex} printing really needs retaining for whatever obscure
> > reason) merely require the processing of %XV to do nothing for EVEX-
> > encoded insns, plus of course the separating blank would then also need
> > to be included in the processing of %XV.
> >
> > I guess I'll make a patch to fold the AVX-VNNI and AVX512-VNNI entries,
> > which you could then re-base on top of.
>
> Folded the table of AVX512IFMA and AVX-IFMA.
>
> >
> > > --- a/opcodes/i386-gen.c
> > > +++ b/opcodes/i386-gen.c
> > > @@ -245,6 +245,8 @@ static initializer cpu_flag_init[] =
> > > "CPU_AVX512F_FLAGS|CpuAVX512_BF16" },
> > > { "CPU_AVX512_FP16_FLAGS",
> > > "CPU_AVX512BW_FLAGS|CpuAVX512_FP16" },
> > > + { "CPU_AVX_IFMA_FLAGS",
> > > + "CPU_AVX2_FLAGS|CpuAVX_IFMA" },
> > > { "CPU_IAMCU_FLAGS",
> > > "Cpu186|Cpu286|Cpu386|Cpu486|Cpu586|CpuIAMCU" },
> > > { "CPU_ADX_FLAGS",
> > > @@ -439,6 +441,8 @@ static initializer cpu_flag_init[] =
> > > "CpuHRESET" },
> > > { "CPU_ANY_AVX512_FP16_FLAGS",
> > > "CpuAVX512_FP16" },
> > > + { "CPU_ANY_AVX_IFMA_FLAGS",
> > > + "CpuAVX_IFMA" },
> >
> > If AVX2 is taken as a prereq feature, then CPU_ANY_AVX2_FLAGS also needs
> > adjustment, such that disabling of AVX2 also results in disabling of
> > AVX-IFMA. (The same issue actually exists for AVX-VNNI afaics.)
> >
>
> Added AVX-IFMA to CPU_ANY_AVX2_FLAGS.
>
> > > --- a/opcodes/i386-opc.tbl
> > > +++ b/opcodes/i386-opc.tbl
> > > @@ -3263,3 +3263,10 @@ vrsqrtph, 0x664e, None, CpuAVX512_FP16,
> > Modrm|Masking=3|EVexMap6|VexW0|Broadcast
> > > vrsqrtsh, 0x664f, None, CpuAVX512_FP16,
> > Modrm|EVexLIG|Masking=3|EVexMap6|VexVVVV|VexW0|Disp8MemShift=1|N
> > o_bSuf|No_wSuf|No_lSuf|No_sSuf|No_qSuf|No_ldSuf,
> > { RegXMM|Word|Unspecified|BaseIndex, RegXMM, RegXMM }
> > >
> > > // FP16 (HFNI) instructions end.
> > > +
> > > +// AVX_IFMA instructions.
> >
> > Nit: Perhaps better use AVX-IFMA here, but I see we're having many examples
> > of the (needless) use of underscores like this.
> >
> > > +vpmadd52huq, 0x66B5, None, CpuAVX_IFMA,
> > Modrm|Vex|PseudoVexPrefix|Space0F38|VexVVVV=1|VexW1|CheckRegSize|N
> > o_bSuf|No_wSuf|No_lSuf|No_sSuf|No_qSuf|No_ldSuf,
> > { RegXMM|RegYMM|Unspecified|BaseIndex, RegXMM|RegYMM,
> > RegXMM|RegYMM }
> > > +vpmadd52luq, 0x66B4, None, CpuAVX_IFMA,
> > Modrm|Vex|PseudoVexPrefix|Space0F38|VexVVVV=1|VexW1|CheckRegSize|N
> > o_bSuf|No_wSuf|No_lSuf|No_sSuf|No_qSuf|No_ldSuf,
> > { RegXMM|RegYMM|Unspecified|BaseIndex, RegXMM|RegYMM,
> > RegXMM|RegYMM }
> >
> > Please use plain VexVVVV (without =1) - we want to have as little clutter as
> > possible on these usually already overlong lines.
>
> Changed to VexVVVV.
>
> Thx for your review and see if there is still something need to be changed.
OK.
Thanks.
> Haochen
>
> >
> > Jan
--
H.J.
More information about the Binutils
mailing list