[PATCH] x86: Determine vector length from the last vector operand

[Jan Beulich]

>>> On 22.07.18 at 21:02, <hjl.tools@gmail.com> wrote:
> Determine VEX/EVEXE vector length from the last multi-length vector
> operand.

I see you've committed this already: It would have been really
helpful to say _why_ you're doing the change in the commit message.
For posterity as well as my understanding - could you at least do so
here please? That's even more so that VEX and EVEX processing
differed in that regard before your change.

> --- a/gas/config/tc-i386.c
> +++ b/gas/config/tc-i386.c
> @@ -3363,10 +3363,12 @@ build_vex_prefix (const insn_template *t)
>      vector_length = 1;
>    else
>      {
> -      unsigned int op;
> +      int op;

This is the sort of change I would have objected to, btw. Variables
used to index arrays should be unsigned whenever possible. And
doing so would have been easy enough here:

> @@ -3611,20 +3613,31 @@ build_evex_prefix (void)
>        if (!i.tm.opcode_modifier.evex
>  	  || i.tm.opcode_modifier.evex == EVEXDYN)
>  	{
> -	  unsigned int op;
> +	  int op;
>  
> +	  /* Determine vector length from the last multi-length vector
> +	     operand.  */
>  	  vec_length = 0;
> -	  for (op = 0; op < i.tm.operands; ++op)
> +	  for (op = i.operands - 1; op >= 0; op--)

	  for (op = i.operands; op--; )

(similarly further down).

>  	    if (i.tm.operand_types[op].bitfield.xmmword
>  		+ i.tm.operand_types[op].bitfield.ymmword
>  		+ i.tm.operand_types[op].bitfield.zmmword > 1)
>  	      {
>  		if (i.types[op].bitfield.zmmword)
> -		  i.tm.opcode_modifier.evex = EVEX512;
> +		  {
> +		    i.tm.opcode_modifier.evex = EVEX512;
> +		    break;
> +		  }
>  		else if (i.types[op].bitfield.ymmword)
> -		  i.tm.opcode_modifier.evex = EVEX256;
> +		  {
> +		    i.tm.opcode_modifier.evex = EVEX256;
> +		    break;
> +		  }
>  		else if (i.types[op].bitfield.xmmword)
> -		  i.tm.opcode_modifier.evex = EVEX128;
> +		  {
> +		    i.tm.opcode_modifier.evex = EVEX128;
> +		    break;
> +		  }
>  		else if (i.broadcast && (int) op == i.broadcast->operand)

This has rendered all the "else" pointless.

Jan