[PATCH] x86: Update register operand check for AddrPrefixOpReg

[Jan Beulich]

On 01.10.2020 15:07, H.J. Lu wrote:
> On Thu, Oct 1, 2020 at 4:35 AM H.J. Lu <hjl.tools@gmail.com> wrote:
>>
>> On Thu, Oct 1, 2020 at 1:22 AM Jan Beulich <jbeulich@suse.com> wrote:
>>>
>>> On 01.10.2020 01:30, H.J. Lu via Binutils wrote:
>>>> I am checking in this patch and backporting it to 2.35 branch.
>>>
>>> But this is wrong, as can be seen from e.g. ...
>>>> + +[a-f0-9]+: 66 0f 38 f8 0d 00 00 00 00      movdir64b 0x0\(%rip\),%rcx        #.*
>>>> + +[a-f0-9]+: 67 66 0f 38 f8 0d 00 00 00 00   movdir64b 0x0\(%eip\),%ecx        #.*
>>>> + +[a-f0-9]+: 67 66 0f 38 f8 0d 00 00 00 00   movdir64b 0x0\(%eip\),%ecx        #.*
>>>
>>> ... the middle line here not matching up with
>>>
>>>> +     movdir64b foo(%rip),%rcx
>>>> +     movdir64b foo(%rip),%ecx
>>>> +     movdir64b foo(%eip),%ecx
>>>
>>> ... what was written here. Without your change this
>>>
>>>         movdir64b (%rbp), %rax
>>>         movdir64b (%rbp), %eax
>>>         movdir64b (%ebp), %rax
>>>         movdir64b (%ebp), %eax
>>>
>>>         movdir64b (%rip), %rax
>>>         movdir64b (%rip), %eax
>>>         movdir64b (%eip), %rax
>>>         movdir64b (%eip), %eax
>>>
>>> yields consistent results for both blocks - the middle two entries
>>> get an error issued.
>>>
>>> Please revert, and once again please don't commit (let alone
>>> backport) patches without giving people at least _a little bit_ of
>>> time to look at them.
>>>
>>
>> The fix is correct.   This specific case came from
>>
>> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97257
>>
>> The address prefix changes the register operand in these instructions.
>> (%rip) is a special case.
>>
> 
> Update the fix for DISP and SYMBOL.

The description wrongly says:

"When the address size prefix applies to both the memory and the register
 operand, we need to extract the address size prefix from the register
 operand if the memory operand has no real registers, like symbol, DISP
 or symbol(%rip)."

And then similarly a code comment. To the assembler, %rip _is_ a real
register (and %eip could be used in its stead, when 32-bit addressing is
desired).

Again - the previous change was wrong and needs undoing. I agree (largely,
not fully anymore, despite my earlier indication of agreement) the plain
symbol/disp case needs fixing; I'm unconvinced though that this is the
place where the fix is to be made (but I didn't invest any time in looking
for alternatives). As a general observation, such extended conditionals to
address isolated cases are often, not always, a good sign that they're
sitting in the wrong place.

As an aside - I wonder how gcc gets away with not using foo(%eip) for
-mx32. But I assume this is due to what I'd call overly heavy restrictions
in the x32 ABI spec.

Jan