Re: [PATCH 2/6] x86/MPX: fix address size handling

["H.J. Lu" <hjl dot tools at gmail dot com>]

On Thu, Oct 10, 2013 at 5:27 AM, Jan Beulich <JBeulich@suse.com> wrote:
>>>> On 09.10.13 at 17:45, "H.J. Lu" <hjl.tools@gmail.com> wrote:
>> On Wed, Oct 9, 2013 at 12:30 AM, Jan Beulich <JBeulich@suse.com> wrote:
>>>>>> On 08.10.13 at 17:32, "H.J. Lu" <hjl.tools@gmail.com> wrote:
>>>> On Tue, Oct 8, 2013 at 8:20 AM, Jan Beulich <JBeulich@suse.com> wrote:
>>>>>>>> On 08.10.13 at 17:15, "H.J. Lu" <hjl.tools@gmail.com> wrote:
>>>>>> On Tue, Oct 8, 2013 at 7:41 AM, Jan Beulich <JBeulich@suse.com> wrote:
>>>>>>> While address overrides are ignored in 64-bit mode (and hence shouldn't
>>>>>>> result in an error), trying to use 16-bit addressing is documented to
>>>>>>> result in #UD, and hence the assembler should reject the attempt.
>>>>>>>
>>>>>>> gas/
>>>>>>> 2013-10-08  Jan Beulich <jbeulich@suse.com>
>>>>>>>
>>>>>>>         * tc-i386.c (md_assemble): Alter address size checking for MPX
>>>>>>>         instructions.
>>>>>>>
>>>>>>> --- 2013-10-07/gas/config/tc-i386.c
>>>>>>> +++ 2013-10-07/gas/config/tc-i386.c
>>>>>>> @@ -3549,10 +3549,15 @@ md_assemble (char *line)
>>>>>>>    if (i.bnd_prefix && !i.tm.opcode_modifier.bndprefixok)
>>>>>>>      as_bad (_("expecting valid branch instruction after `bnd'"));
>>>>>>>
>>>>>>> -  if (i.tm.cpu_flags.bitfield.cpumpx
>>>>>>> -      && flag_code == CODE_64BIT
>>>>>>> -      && i.prefix[ADDR_PREFIX])
>>>>>>> -    as_bad (_("32-bit address isn't allowed in 64-bit MPX instructions."));
>>>>>>
>>>>>> It is done on purpose.  When 32-bit address prefix in 64-bit is ignored,
>>>>>> MPX doesn't work correctly for x32.
>>>>>
>>>>> I don't understand: It _is_ being ignored by the hardware as per
>>>>> the documentation. So x32 need to get along with that. Maybe
>>>>> an example would help, so I could understand why you think
>>>>> this _needs_ to be an error...
>>>>>
>>>>
>>>> X32 won't work with MPX since hardware assumes pointer
>>>> size is always 64 bit in 64-bit mode with or without address
>>>> size prefix.  MPX depends on correct pointer size to work.
>>>> I don't want people to use MPX in x32 by accident.
>>>
>>> This seems even more odd - why would x32 be excluded from
>>> using MPX? Again - an example might help, as my understanding
>>> so far was that the implicit zero extension of results of 32-bit
>>> operations should guarantee the half width pointers to be quite
>>> fine to use as full width values (i.e. in other memory operands
>>> I don't see why you would want to use 32-bit addressing either,
>>> except when the wraparound case matters, as might e.g. be
>>> the case with EIP-relative addressing).
>>
>> In 64-bit mode, bndldx and bndstx ignore the lower 3 bits of
>> the address of a pointer, which is OK when pointers are
>> 64-bit aligned. X32 runs in 64-bit mode. But pointers are
>> 32-bit aligned.  That means 2 pointers may point to the same
>> bound table entry.  That is why MPX won't work for x32 and
>> assembler shouldn't allow it.
>
> Here you're making assumptions that you can't control. For
> example, there's nothing preventing anyone to create a
> compiler/library pair that guarantees objects to be at least 8
> bytes apart. And that's leaving aside that even in full 64-bit
> mode there can be multiple objects within an 8-byte range.
>

It won't work with

struct foo
{
  void *x;
  void *y;
  void *z;
  void *p;
};

nor

int *foo[20];

2 pointers will share the same bound table entry.

-- 
H.J.