Re: [PATCH] x86/Intel: don't accept bogus instructions

["Jan Beulich" <JBeulich at suse dot com>]

>>> On 30.06.16 at 14:07, <hjl.tools@gmail.com> wrote:
> On Thu, Jun 30, 2016 at 4:58 AM, Jan Beulich <JBeulich@suse.com> wrote:
>>>>> On 30.06.16 at 13:50, <hjl.tools@gmail.com> wrote:
>>> On Thu, Jun 30, 2016 at 4:38 AM, Jan Beulich <JBeulich@suse.com> wrote:
>>>>>>> On 30.06.16 at 13:18, <hjl.tools@gmail.com> wrote:
>>>>> On Thu, Jun 30, 2016 at 3:34 AM, Jan Beulich <JBeulich@suse.com> wrote:
>>>>>> ... due to their last byte looking like a suffix, when after its
>>>>>> stripping a matching instruction can be found. Since memory operand
>>>>>> size specifiers in Intel mode get converted into suffix representation
>>>>>> internally, we need to keep track of the actual mnemonic suffix which
>>>>>> may have got trimmed off, and check its validity while looking for a
>>>>>> matching template. I tripper over this quite some time again after
>>>>>> support for AMD's SSE5 instructions got removed, as at that point some
>>>>>> of the SSE5 mnemonics, other than expected, didn't fail to assemble.
>>>>>> But the problem affects many more instructions, namely (almost) all
>>>>>> MMX, SSE, and AVX ones as it looks. I don't think it makes sense to
>>>>>> add a testcase covering all of them, nor do I think it makes sense to
>>>>>> pick out some random examples for a new test case.
>>>>>
>>>>> Please open a bug report to show there is a problem.
>>>>
>>>> I don't see the point, but anyway: 20318.
>>>
>>> Please add this testcase:
>>>
>>> .text
>>> phadddb %xmm0, %xmm1
>>> phadddd %xmm0, %xmm1
>>> phadddl %xmm0, %xmm1
>>> phadddld %xmm0, %xmm1
>>> phadddq %xmm0, %xmm1
>>> phaddds %xmm0, %xmm1
>>> phadddw %xmm0, %xmm1
>>> .intel_syntax noprefix
>>> phadddb xmm0, xmm1
>>> phadddd xmm0, xmm1
>>> phadddl xmm0, xmm1
>>> phadddld xmm0, xmm1
>>> phadddq xmm0, xmm1
>>> phaddds xmm0, xmm1
>>> phadddw xmm0, xmm1
>>>
>>> to your patch and run it with both --32 and --64.  All of them should fail.
>>
>> To quote the original submission: "I don't think it makes sense to add
>> a testcase covering all of them, nor do I think it makes sense to pick
>> out some random examples for a new test case."
>>
>> What good does such a limited test case do? If anything, someone
>> should sit down and write a script to machine generate a list of all
>> mnemonics with invalid suffixes.
> 
> Without a testcase, your change looks like random change to me.

Sigh.

> For all I know, you change doesn't make those instructions error.

Of course it does - that's the whole point.

Jan