RE: [PATCH 7/8] Support APX NF

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

> Subject: Re: [PATCH 7/8] Support APX NF
> 
> On 19.09.2023 17:25, Cui, Lili wrote:
> > @@ -4178,11 +4180,15 @@ build_evex_insns_with_extend_evex_prefix
> (void)
> >      i.vex.bytes[1] &= 0xef;
> >    if (i.vex.register_specifier
> >        && register_number (i.vex.register_specifier) > 0xf)
> > -    i.vex.bytes[3] &=0xf7;
> > +    i.vex.bytes[3] &= 0xf7;
> 
> When you notice such issues, they want correcting in the patch introducing
> them.
> 

Fixed it in the original patch.

> > @@ -5944,6 +5954,10 @@ parse_insn (const char *line, char *mnemonic,
> bool prefix_only)
> >  		  /* {rex2} */
> >  		  i.rex2_encoding = true;
> >  		  break;
> > +		case Prefix_NF:
> > +		  /* {NF} */
> > +		  i.has_nf = true;
> > +		  break;
> 
> I find it odd that this is represented as a (pseudo-)prefix. The manual doesn't
> suggest so; it rather looks like the intention is for it to be a mnemonic suffix,
> as in "add{nf} ...". Hence same question as before: In how far is this
> representation aligned with what other assemblers are going to do?
> 

Jan, thanks for the great suggestions on these prefixes and suffixes. We will consider each item carefully and try to document these special things. This patch needs to be suspended.

> > @@ -7151,6 +7165,19 @@ optimize_NDD_to_nonNDD (const insn_template
> *t)
> 
> How useful that this function is mentioned at least this way: No change there?
> (See my comments on the patch introducing it.)
> 
 
Reserved has_nf flag in a previous patch.

> >    return 0;
> >  }
> >
> > +/* Check if NF prefix requirements are met by the instruction.  */
> > +static int
> 
> As before, bool please for functions returning boolean values.
> 
Done.

> > +check_NfPrefix (const insn_template *t) {
> > +  if (i.has_nf && !t->opcode_modifier.nf)
> > +    {
> > +      /* This instruction should support nf prefix.  */
> > +      i.error = unsupported;
> 
> A more specific error message would be nice here.
> 
Done.

> Question of course is whether, for such an isolated check, you really need a
> new helper function.
> 
Removed the function.

> > @@ -7551,6 +7578,7 @@ match_template (char mnem_suffix)
> >  		  goto check_operands_345;
> >  		}
> >  	      else if (t->opcode_space != SPACE_BASE
> > +		       && !t->opcode_modifier.nf
> >  		       && (t->opcode_space != SPACE_0F
> >  			   /* MOV to/from CR/DR/TR, as an exception, follow
> >  			      the base opcode space encoding model.  */
> 
> With an earlier comment addressed, I expect this change may not be
> necessary anymore.
> 
Yes, it has been removed.

> > @@ -7652,6 +7680,13 @@ match_template (char mnem_suffix)
> >  	  continue;
> >  	}
> >
> > +      /* Check if nf prefix are valid.  */
> > +      if (check_NfPrefix (t))
> > +	{
> > +	  specific_error = progress (i.error);
> > +	  continue;
> > +	}
> 
> Is it helpful (e.g. diagnostic-wise) to have this check so late? If so, is it useful to
> "continue" when this is the only thing that doesn't match?
> No other template is going to match in such an event, afaict.
> 

Yes, it is useful to continue.
For example" {nf}  add  $123, %bl "
In order, we first find legacy add, then the judgment finds they don’t match, then continue to find the second template which has NF.

add, 0x83/0, 0, Modrm|No_bSuf|No_sSuf|HLEPrefixLock, { Imm8S, Reg16|Reg32|Reg64|Word|Dword|Qword|Unspecified|BaseIndex }
add, 0x83/0, APX_F, Modrm|No_bSuf|No_sSuf|EVex128|EVexMap4|NF, { Imm8S, Reg16|Reg32|Reg64|Word|Dword|Qword|Unspecified|BaseIndex }

> > --- a/gas/testsuite/gas/i386/x86-64-apx-ndd.d
> > +++ b/gas/testsuite/gas/i386/x86-64-apx-ndd.d
> > @@ -158,7 +158,7 @@ Disassembly of section .text:
> >  \s*[a-f0-9]+:\s*67 62 f4 3c 18 4f 90 90 90 90 90 	cmovg  -
> 0x6f6f6f70\(%eax\),%edx,%r8d
> >  \s*[a-f0-9]+:\s*67 62 f4 3c 18 af 90 09 09 09 00 	imul
> 0x90909\(%eax\),%edx,%r8d
> >  \s*[a-f0-9]+:\s*62 b4 b0 10 af 94 f8 09 09 00 00 	imul
> 0x909\(%rax,%r31,8\),%rdx,%r25
> > -\s*[a-f0-9]+:\s*62 f4 fc 08 ff c0\s+inc    %rax
> > +\s*[a-f0-9]+:\s*62 f4 fc 08 ff c0\s+\{evex\} inc %rax
> 
> It's kind of unexpected to see this change here.
>
It has been removed in the NDD patch.

> > --- /dev/null
> > +++ b/gas/testsuite/gas/i386/x86-64-apx-nf.s
> > @@ -0,0 +1,1256 @@
> > +# Check 64bit APX_F instructions
> > +
> > +        .text
> > +_start:
> > +	{nf}	add	$123, %bl	 #APX_F OPC_EVEX_NF
> OPC_EVEX_EVEX
> 
> Comments on earlier patches apply throughout this file as well.
> 

Done.

> > +	{nf}	imul	291(%r8, %rax, 4), %ecx, %edx	 #APX_F
> OPC_EVEX_NF OPC_EVEX_ND
> > +	{nf}	imulq	291(%r8, %rax, 4)	 #APX_F OPC_EVEX_NF
> OPC_EVEX_EVEX
> > +	{nf}	imul	291(%r8, %rax, 4), %r9	 #APX_F OPC_EVEX_NF
> OPC_EVEX_EVEX
> > +	{nf}	imul	291(%r8, %rax, 4), %r9, %r31	 #APX_F
> OPC_EVEX_NF OPC_EVEX_ND
> 
> No IMUL by immediate?
> 

Replied with the earlier patch.

> > --- a/gas/testsuite/gas/i386/x86-64.exp
> > +++ b/gas/testsuite/gas/i386/x86-64.exp
> > @@ -372,6 +372,8 @@ run_dump_test "x86-64-apx-evex-promoted"
> >  run_dump_test "x86-64-apx-evex-promoted-intel"
> >  run_dump_test "x86-64-apx-evex-egpr"
> >  run_dump_test "x86-64-apx-ndd"
> > +run_dump_test "x86-64-apx-nf"
> > +run_dump_test "x86-64-apx-nf-intel"
> >  run_dump_test "x86-64-avx512f-rcigrz-intel"
> >  run_dump_test "x86-64-avx512f-rcigrz"
> >  run_dump_test "x86-64-clwb"
> 
> No test checking that {nf} isn't accepted (assembler) / EVEX.nf set is rejected
> (disassembler) on insns not permitting its use, at least for a few examples?
> 
Done.

> > @@ -1003,7 +1007,7 @@ typedef struct insn_template
> >       AMD 3DNow! instructions.
> >       If this template has no extension opcode (the usual case) use None
> >       Instructions */
> > -  signed int extension_opcode:0xA;
> > +  signed int extension_opcode:0xB;
> 
> For this and ...
> 
> > @@ -1017,7 +1021,8 @@ typedef struct insn_template
> >  #define Prefix_EVEX		7	/* {evex} */
> >  #define Prefix_REX		8	/* {rex} */
> >  #define Prefix_REX2		9	/* {rex2} */
> > -#define Prefix_NoOptimize	0xA	/* {nooptimize} */
> > +#define Prefix_NF		0xA	/* {nf} */
> > +#define Prefix_NoOptimize	0xB	/* {nooptimize} */
> 
> ... this, see comments on an earlier patch.
> 

Done.

> > --- a/opcodes/i386-opc.tbl
> > +++ b/opcodes/i386-opc.tbl
> > @@ -286,25 +286,41 @@ add, 0x0, 0,
> > D|W|CheckOperandSize|Modrm|No_sSuf|HLEPrefixLock, { Reg8|Reg16|Reg3
> > add, 0x83/0, 0, Modrm|No_bSuf|No_sSuf|HLEPrefixLock, { Imm8S,
> > Reg16|Reg32|Reg64|Word|Dword|Qword|Unspecified|BaseIndex }  add,
> 0x4,
> > 0, W|No_sSuf, { Imm8|Imm16|Imm32|Imm32S,
> Acc|Byte|Word|Dword|Qword }
> > add, 0x80/0, 0, W|Modrm|No_sSuf|HLEPrefixLock, {
> > Imm8|Imm16|Imm32|Imm32S,
> >
> Reg8|Reg16|Reg32|Reg64|Byte|Word|Dword|Qword|Unspecified|BaseIndex
> }
> > -add, 0x0, APX_F|x64,
> > D|W|CheckOperandSize|Modrm|No_sSuf|VexVVVV|EVex128|EVexMap4, {
> > Reg8|Reg16|Reg32|Reg64,
> >
> Reg8|Reg16|Reg32|Reg64|Byte|Word|Dword|Qword|Unspecified|BaseIndex,
> > Reg8|Reg16|Reg32|Reg64 } -add, 0x83/0, APX_F|x64,
> >
> Modrm|CheckOperandSize|No_bSuf|No_sSuf|VexVVVV|EVex128|EVexMap4, {
> > Imm8S, Reg16|Reg32|Reg64|Word|Dword|Qword|Unspecified|BaseIndex,
> > Reg16|Reg32|Reg64 } -add, 0x80/0, APX_F|x64,
> > W|Modrm|CheckOperandSize|No_sSuf|VexVVVV|EVex128|EVexMap4, {
> > Imm8|Imm16|Imm32|Imm32S,
> >
> Reg8|Reg16|Reg32|Reg64|Byte|Word|Dword|Qword|Unspecified|BaseIndex,
> > Reg8|Reg16|Reg32|Reg64}
> > +
> > +add, 0x0, APX_F|x64,
> > +D|W|CheckOperandSize|Modrm|No_sSuf|EVex128|EVexMap4|NF, {
> > +Reg8|Reg16|Reg32|Reg64,
> >
> +Reg8|Reg16|Reg32|Reg64|Byte|Word|Dword|Qword|Unspecified|BaseInde
> x }
> > +add, 0x83/0, APX_F|x64,
> Modrm|No_bSuf|No_sSuf|EVex128|EVexMap4|NF, {
> > +Imm8S, Reg16|Reg32|Reg64|Word|Dword|Qword|Unspecified|BaseIndex }
> > +add, 0x80/0, APX_F|x64, W|Modrm|No_sSuf|EVex128|EVexMap4|NF, {
> > +Imm8|Imm16|Imm32|Imm32S,
> >
> +Reg8|Reg16|Reg32|Reg64|Byte|Word|Dword|Qword|Unspecified|BaseInde
> x }
> 
> Huge patches like this are already hard enough to review. Can you please try
> to make sure you introduce new templates right in their final shape (within
> the specific series of course), rather than touching them again a 2nd time?
> Even without fully supporting NF, introducing the attribute (as a dummy or
> without any consumer) ought to be possible earlier on.
> 
Added NF in NDD patch.

> As per earlier comments many of these templates need cleaning up anyway, so
> I won't look at the other in any detail here, and instead wait for a v2.
> 
Ok.

Thanks,
Lili.