[PATCH 1/2] Add check for 8-bit old registers in EVEX format

[Jan Beulich]

On 17.05.2024 03:45, Cui, Lili wrote:
>> On 15.05.2024 08:31, Cui, Lili wrote:
>>> --- a/gas/config/tc-i386.c
>>> +++ b/gas/config/tc-i386.c
>>> @@ -7029,6 +7029,18 @@ md_assemble (char *line)
>>>  	  as_bad (_("{rex2} prefix invalid with `%s'"), insn_name (&i.tm));
>>>  	  return;
>>>  	}
>>> +      /* Check for 8 bit operand that uses old registers.  */
>>> +      for (unsigned int op = 0; op < i.operands; op++)
>>> +	{
>>> +	  if (i.types[op].bitfield.class == Reg
>>> +	      && i.types[op].bitfield.byte
>>> +	      && !(i.op[op].regs->reg_flags & RegRex64)
>>> +	      && i.op[op].regs->reg_num > 3)
>>> +
>>> +	    as_bad (_("can't encode register '%s' in an "
>>> +		      " EVEX/VEX prefix instruction"),
>>> +		    i.op[op].regs->reg_name);
>>> +	}
>>
>> I'd like us to avoid duplicating the conditional used here, matching what
>> establish_rex() does. establish_rex() is called unconditionally, so there are two
>> questions (both of which could have been addressed if there wasn't, once
>> again, an entirely empty patch description here): Why does the checking there
>> not cover this case? Is it not possible to amend that logic, rather than
>> introducing another instance in an entirely distinct place?
>>
> 
> Agreed, when I added this patch, I first tried adding it with rex/rex2. But since that judgment of establish_rex() is tied to "i.rex", as you also suggested, I created a patch to detach the old register check from that branch and then extended it to check the EVEX prefixed instructions .

Yet that tying to "i.rex" could be amended, I think.

Jan