[PATCH 0/2] x86: Fix the -mevexwig=1 assembler option

[H.J. Lu]

On Sun, Sep 16, 2018 at 1:48 AM, Jan Beulich <jbeulich@suse.com> wrote:
>>>> "H.J. Lu" <hjl.tools@gmail.com> 09/14/18 9:14 PM >>>
>>The VEX.W/EVEX.W bit is ignored by some VEX/EVEX instructions, aka WIG
>>instructions.  The -mevexwig=1 assembler option assumes that if the
>>vexw field of an VEX/EVEX instruction is 0, it is a WIG instruction.
>>But the vexw field of some non-WIG VEX/EVEX instructions is 0 and their
>>VEX.W/EVEX.W bit is determined by the integer register operand size.  This
>>patch set adds VEXWIG, defined as 3, to indicate that the VEX.W/EVEX.W
>>bit is ignored and set VexW=3 on VEX/EVEX WIG instructions.
>
> Now this is exactly what I was afraid of, and what should imo absolutely not
> happen anymore going forward: You introduce yet another disconnect between
> i386-opc.h and i386-opc.tbl - one can't freely change the VEXWIG define without
> having to touch a large number of templates. The .tbl file gets pre-processed now,
> so all new attributes of non-boolean type should get added as (taking the
> example here) VexWIG in the templates, with
>
> #define VexWIG VexW=VEXWIG
>
> somewhere earlier in the file. It is going to be yet another time consuming job
> to convert all the existing instances, to correct the mistakes made here years
> ago. _Then_, finally, it will be half way reasonable to read and maintain all those
> templates.

A good point.  I will take a look.

-- 
H.J.