[PATCH 4/8] Support APX NDD

Cui, Lili lili.cui@intel.com
Tue Oct 24 11:39:57 GMT 2023


> Subject: Re: [PATCH 4/8] Support APX NDD
> 
> On 19.09.2023 17:25, Cui, Lili wrote:
> > --- a/opcodes/i386-dis-evex-prefix.h
> > +++ b/opcodes/i386-dis-evex-prefix.h
> > @@ -353,8 +353,8 @@
> >    /* PREFIX_EVEX_MAP4_66 */
> >    {
> >      { MOD_TABLE (MOD_EVEX_MAP4_66_PREFIX_0) },
> > -    { "adoxS",	{ Gdq, Edq }, 0 },
> > -    { "adcxS",	{ Gdq, Edq }, 0 },
> > +    { "adoxS",	{ VexGdq, Gdq, Edq }, 0 },
> > +    { "adcxS",	{ VexGdq, Gdq, Edq }, 0 },
> 
> With the OP_VEX() change moved to the earlier patch, you wouldn't need to
> alter again right away what was just added there.
> 

As I commented in another email, I want to move instructions like adcx into this NDD patch. Do you think it is ok?

> > --- a/opcodes/i386-dis-evex-reg.h
> > +++ b/opcodes/i386-dis-evex-reg.h
> > @@ -56,6 +56,105 @@
> >      { "blsmskS",	{ VexGdq, Edq }, 0 },
> >      { "blsiS",		{ VexGdq, Edq }, 0 },
> >    },
> > +  /* REG_EVEX_MAP4_80 */
> > +  {
> > +    { "addA",	{ VexGb, Eb, Ib }, 0 },
> > +    { "orA",	{ VexGb, Eb, Ib }, 0 },
> > +    { "adcA",	{ VexGb, Eb, Ib }, 0 },
> > +    { "sbbA",	{ VexGb, Eb, Ib }, 0 },
> 
> Aren't these two and other adc/sbb entries required already in the earlier
> patch, for consistency with what you add there on the assembler side?
> 
> > +    { "andA",	{ VexGb, Eb, Ib }, 0 },
> > +    { "subA",	{ VexGb, Eb, Ib }, 0 },
> > +    { "xorA",	{ VexGb, Eb, Ib }, 0 },
> > +    { Bad_Opcode },
> > +  },
> > +  /* REG_EVEX_MAP4_81 */
> > +  {
> > +    { "addQ",	{ VexGv, Ev, Iv }, 0 },
> > +    { "orQ",	{ VexGv, Ev, Iv }, 0 },
> > +    { "adcQ",	{ VexGv, Ev, Iv }, 0 },
> > +    { "sbbQ",	{ VexGv, Ev, Iv }, 0 },
> > +    { "andQ",	{ VexGv, Ev, Iv }, 0 },
> > +    { "subQ",	{ VexGv, Ev, Iv }, 0 },
> > +    { "xorQ",	{ VexGv, Ev, Iv }, 0 },
> > +    { Bad_Opcode },
> > +  },
> > +  /* REG_EVEX_MAP4_83 */
> > +  {
> > +    { "addQ",	{ VexGv, Ev, sIb }, 0 },
> > +    { "orQ",	{ VexGv, Ev, sIb }, 0 },
> > +    { "adcQ",	{ VexGv, Ev, sIb }, 0 },
> > +    { "sbbQ",	{ VexGv, Ev, sIb }, 0 },
> > +    { "andQ",	{ VexGv, Ev, sIb }, 0 },
> > +    { "subQ",	{ VexGv, Ev, sIb }, 0 },
> > +    { "xorQ",	{ VexGv, Ev, sIb }, 0 },
> > +    { Bad_Opcode },
> > +  },
> 
> While these are needed because of the differences (from existing table entries
> we have) in the last entry (albeit I don't think those last entries need to
> actually be spelled out; you don't spell out trailing invalid entries further
> down, and we don't do so elsewhere either), ...

Done.

> 
> > +  /* REG_EVEX_MAP4_C0 */
> > +  {
> > +    { "rolA",	{ VexGb, Eb, Ib }, 0 },
> > +    { "rorA",	{ VexGb, Eb, Ib }, 0 },
> > +    { "rclA",	{ VexGb, Eb, Ib }, 0 },
> > +    { "rcrA",	{ VexGb, Eb, Ib }, 0 },
> > +    { "shlA",	{ VexGb, Eb, Ib }, 0 },
> > +    { "shrA",	{ VexGb, Eb, Ib }, 0 },
> > +    { "shlA",	{ VexGb, Eb, Ib }, 0 },
> > +    { "sarA",	{ VexGb, Eb, Ib }, 0 },
> > +  },
> > +  /* REG_EVEX_MAP4_C1 */
> > +  {
> > +    { "rolQ",	{ VexGv, Ev, Ib }, 0 },
> > +    { "rorQ",	{ VexGv, Ev, Ib }, 0 },
> > +    { "rclQ",	{ VexGv, Ev, Ib }, 0 },
> > +    { "rcrQ",	{ VexGv, Ev, Ib }, 0 },
> > +    { "shlQ",	{ VexGv, Ev, Ib }, 0 },
> > +    { "shrQ",	{ VexGv, Ev, Ib }, 0 },
> > +    { "shlQ",	{ VexGv, Ev, Ib }, 0 },
> > +    { "sarQ",	{ VexGv, Ev, Ib }, 0 },
> > +  },
> > +  /* REG_EVEX_MAP4_D0 */
> > +  {
> > +    { "rolA",	{ VexGb, Eb, I1 }, 0 },
> > +    { "rorA",	{ VexGb, Eb, I1 }, 0 },
> > +    { "rclA",	{ VexGb, Eb, I1 }, 0 },
> > +    { "rcrA",	{ VexGb, Eb, I1 }, 0 },
> > +    { "shlA",	{ VexGb, Eb, I1 }, 0 },
> > +    { "shrA",	{ VexGb, Eb, I1 }, 0 },
> > +    { "shlA",	{ VexGb, Eb, I1 }, 0 },
> > +    { "sarA",	{ VexGb, Eb, I1 }, 0 },
> > +  },
> > +  /* REG_EVEX_MAP4_D1 */
> > +  {
> > +    { "rolQ",	{ VexGv, Ev, I1 }, 0 },
> > +    { "rorQ",	{ VexGv, Ev, I1 }, 0 },
> > +    { "rclQ",	{ VexGv, Ev, I1 }, 0 },
> > +    { "rcrQ",	{ VexGv, Ev, I1 }, 0 },
> > +    { "shlQ",	{ VexGv, Ev, I1 }, 0 },
> > +    { "shrQ",	{ VexGv, Ev, I1 }, 0 },
> > +    { "shlQ",	{ VexGv, Ev, I1 }, 0 },
> > +    { "sarQ",	{ VexGv, Ev, I1 }, 0 },
> > +  },
> > +  /* REG_EVEX_MAP4_D2 */
> > +  {
> > +    { "rolA",	{ VexGb, Eb, CL }, 0 },
> > +    { "rorA",	{ VexGb, Eb, CL }, 0 },
> > +    { "rclA",	{ VexGb, Eb, CL }, 0 },
> > +    { "rcrA",	{ VexGb, Eb, CL }, 0 },
> > +    { "shlA",	{ VexGb, Eb, CL }, 0 },
> > +    { "shrA",	{ VexGb, Eb, CL }, 0 },
> > +    { "shlA",	{ VexGb, Eb, CL }, 0 },
> > +    { "sarA",	{ VexGb, Eb, CL }, 0 },
> > +  },
> > +  /* REG_EVEX_MAP4_D3 */
> > +  {
> > +    { "rolQ",	{ VexGv, Ev, CL }, 0 },
> > +    { "rorQ",	{ VexGv, Ev, CL }, 0 },
> > +    { "rclQ",	{ VexGv, Ev, CL }, 0 },
> > +    { "rcrQ",	{ VexGv, Ev, CL }, 0 },
> > +    { "shlQ",	{ VexGv, Ev, CL }, 0 },
> > +    { "shrQ",	{ VexGv, Ev, CL }, 0 },
> > +    { "shlQ",	{ VexGv, Ev, CL }, 0 },
> > +    { "sarQ",	{ VexGv, Ev, CL }, 0 },
> > +  },
> 
> ... do we really need all these new entries? OP_VEX() checks need_vex first
> thing, so simply adjusting and then re-using the existing entries would seem
> possible.
> 

Done.

> > @@ -9070,6 +9085,14 @@ get_valid_dis386 (const struct dis386 *dp,
> instr_info *ins)
> >  	  ins->rex &= ~REX_B;
> >  	  ins->rex2 &= ~REX_R;
> >  	}
> > +      if (ins->evex_type == evex_from_legacy)
> > +	{
> > +	  if (ins->vex.ll || ins->vex.zeroing
> > +	      || (!ins->vex.b && (ins->vex.register_specifier
> > +				  || !ins->vex.v)))
> > +	    return &bad_opcode;
> 
> Don't these checks also apply to evex_from_vex?

Oh, I've duplicated some of the checks in the above code and removed the duplicated parts.

     if (ins->evex_type == evex_from_legacy)
        {
          /* EVEX from legacy instructions, when the EVEX.ND bit is 0,
             all bits of EVEX.vvvv and EVEX.V' must be 1.  */
          if (!ins->vex.b && (ins->vex.register_specifier
                                  || !ins->vex.v))
            return &bad_opcode;
          ins->rex |= REX_OPCODE;
        }

      /* EVEX from legacy instructions require that EVEX.L’L, EVEX.z and the
         lower 2 bits of EVEX.aaa must be 0.
         EVEX from evex instrucions require that EVEX.L’L and the lower 2 bits of
         EVEX.aaa must be 0.  */
      if (ins->evex_type == evex_from_legacy || ins->evex_type == evex_from_vex)
        {
          if ((*ins->codep & 0x3) != 0
              || (*ins->codep >> 6 & 0x3) != 0
              || (ins->evex_type == evex_from_legacy
                  && (*ins->codep >> 5 & 0x1) != 0)
              || (ins->evex_type == evex_from_vex
                  && !ins->vex.b))
            return &bad_opcode;
        }
> 
> > +	  ins->rex |= REX_OPCODE;
> 
> This, otoh, may truly be evex_from_legacy-only (but I'm not entirely certain).
>

Added after separation.
 
> > @@ -9080,7 +9103,7 @@ get_valid_dis386 (const struct dis386 *dp,
> instr_info *ins)
> >  	return &err_opcode;
> >
> >        /* Set vector length.  */
> > -      if (ins->modrm.mod == 3 && ins->vex.b)
> > +      if (ins->modrm.mod == 3 && ins->vex.b && ins->evex_type ==
> > + evex_default)
> >  	ins->vex.length = 512;
> 
> Down from here, still in the same function, there's another vex.b check which I
> think also wants qualifying.
> 

Not found in this function, but found a place that needs to be qualified in print_insn.

          /* Check whether rounding control was enabled for an insn not
             supporting it.  */
          if (ins.modrm.mod == 3 && ins.vex.b
              && !(ins.evex_used & EVEX_b_used))
            {
              for (i = 0; i < MAX_OPERANDS; ++i)
                {
                  ins.obufp = ins.op_out[i];
                  if (*ins.obufp)
                    continue;
                  oappend (&ins, names_rounding[ins.vex.ll]);
                  oappend (&ins, "bad}");
                  break;
                }
            }

> > @@ -10994,7 +11017,7 @@ print_displacement (instr_info *ins,
> > bfd_signed_vma val)  static void  intel_operand_size (instr_info *ins,
> > int bytemode, int sizeflag)  {
> > -  if (ins->vex.b)
> > +  if (ins->vex.b && ins->evex_type != evex_from_legacy)
> 
> Wouldn't this better be ins->evex_type == evex_default, ...
> 

Done.

> > @@ -11928,7 +11951,8 @@ OP_E_memory (instr_info *ins, int bytemode,
> int sizeflag)
> >  	ins->vex.no_broadcast = true;
> >
> >        if (!ins->vex.no_broadcast
> > -	  && (!ins->intel_syntax || !(ins->evex_used & EVEX_len_used)))
> > +	  && (!ins->intel_syntax || !(ins->evex_used & EVEX_len_used))
> > +	  && ins->evex_type == evex_default)
> 
> ... just like you have it here?
> 
> However, for this change, doesn't this need moving to the enclosing if()?
> You should accidentally set EVEX_b_used here for APX insns.
> 

Sure, moved it.

> > @@ -13280,6 +13304,14 @@ OP_VEX (instr_info *ins, int bytemode, int
> sizeflag ATTRIBUTE_UNUSED)
> >    if (!ins->need_vex)
> >      return true;
> >
> > +  if (ins->evex_type == evex_from_legacy)
> > +    {
> > +      if (ins->vex.b)
> > +	ins->evex_used |= EVEX_b_used;
> > +      else
> > +	 return true;
> > +    }
> 
> When you reuse fields or definitions in places where their names don't match
> their purpose (the field dealt with here is "nd" after all, not "b"), a comment
> wants adding. There's also something odd with indentation here, but I
> suppose an if/else construct isn't needed in the first place.
> 
Added comment and adjusted the formatting, I think we need if/else like "adox " which supports EVEX.ND=0/1.

/* PREFIX_0F38F6 */
  {
    { "wrssK",  { M, Gdq }, 0 },
    { "adoxS",  { VexGdq, Gdq, Edq}, 0 },

adox, 0xf366, ADX|APX_F, Modrm|CheckOperandSize|IgnoreSize|No_bSuf|No_wSuf|No_sSuf|EVex128|EVexMap4, { Reg32|Reg64|Unspecified|BaseIndex, Reg32|Reg64 }
adox, 0xf366, ADX|APX_F, Modrm|CheckOperandSize|IgnoreSize|No_bSuf|No_wSuf|No_sSuf|VexVVVVDest|EVex128|EVexMap4, { Reg32|Reg64|Unspecified|BaseIndex, Reg32|Reg64, Reg32|Reg64 }


Thanks,
Lili.




More information about the Binutils mailing list