Re: [PATCH, v2] gas/x86: don't allow invalid operand combinations for VGATHER

["Jan Beulich" <JBeulich@suse.com>]

>>> On 07.08.12 at 15:21, "H.J. Lu" <hjl.tools@gmail.com> wrote:
> On Tue, Aug 7, 2012 at 3:28 AM, Jan Beulich <JBeulich@suse.com> wrote:
>>>>> On 31.07.12 at 17:43, "H.J. Lu" <hjl.tools@gmail.com> wrote:
>>> On Tue, Jul 31, 2012 at 12:49 AM, Jan Beulich <JBeulich@suse.com> wrote:
>>>>>>> On 30.07.12 at 18:10, "H.J. Lu" <hjl.tools@gmail.com> wrote:
>>>>> The assembler should only check the operands which can't be
>>>>> encoded.  It should shouldn't check if operands are functional
>>>>> correct.  However, I don't mind to issue an error which is controlled
>>>>> by a command line option.
>>>>
>>>> Hmm, not sure. Is there any precedent to such behavior? I as a
>>>> programmer would appreciate if the assembler rejected anything
>>>> that's invalid.
>>>>
>>>> In the case you stay on that position, would making the new
>>>> diagnostic an unconditional warning be acceptable instead?
>>>>
>>>
>>> Can you also add a command line option to turn it off and
>>> turn warning into error?
>>
>> How about this then?
>>
>> Jan
>>
>> The VGATHER group of instructions requires that all three involved
>> xmm/ymm registers are distinct. This patch adds code to check for this,
>> and at once eliminates a superfluous check for not using PC-relative
>> addressing for these instructions (the fact that an index register is
>> required here already excludes valid PC-relative addresses). The
>> severity of the resulting diagnostics can be controlled via command
>> line option or directive.
> 
> Why
> 
> vgatherdps %xmm2,(%rax,%xmm1,2),%xmm2
> 
> can't be used?  XMM2 can be used for both mask and
> destination.

Not that I'm aware of (quoting the documentation):

"If any pair of the index, mask, or destination registers are the
 same, this instruction results a #UD fault."

It's my understanding that this is because of the restartability of
the instruction (i.e. if the same register served as mask and
destination, the instruction might never complete if the loaded
vector element had its high bit set).

Jan