Untested code

[Rafael Espíndola via llvm-commits]

On 13 April 2016 at 16:50, Simon Atanasyan <simon at atanasyan.com> wrote:
> On Wed, Apr 13, 2016 at 11:02 PM, Rafael Espíndola
> <rafael.espindola at gmail.com> wrote:
>> On 13 April 2016 at 10:58, Simon Atanasyan <simon at atanasyan.com> wrote:
>>> On Wed, Apr 13, 2016 at 12:02 AM, Rafael Espíndola
>>> <rafael.espindola at gmail.com> wrote:
>>>> I think now there is only one if left untested (see the attached patch) :-)
>>>
>>> Attached patch adds test case and tries to fix the problem.
>>>
>>> _gp_disp designates offset between start of function and 'gp' pointer into GOT.
>>> Usually you can see the following code in the beginning of a function:
>>>
>>> lui    $t0, %hi(_gp_disp)
>>> addi   $t0, $t0, %lo(_gp_disp)
>>>
>>> When you calculate result of %hi(_gp_disp), you use %hi(_gp - P)
>>> expression. For the first instruction "_gp - P" gives the distance
>>> between the function and _gp. But for the second instruction you have
>>> to add four byte offset %hi(_gp - P + 4). Code in the trunk
>>> incorrectly calculates %lo(_gp_disp) as an absolute expression
>>> %lo(_gp).
>>>
>>> I am not sure that the fix is ideal. So I would appreciate any
>>> suggestions / objections.
>>
>> Is it guaranteed that _gp_disp can only ever be used as the first two
>> instructions in a function? Is that restriction documented somewhere?
>> If so, please add a link to the documentation.
>
> Yes. This pair of instructions is a standard function preamble. I will
> try to find a link to spec and add it to the code.
>
>> Instead of changing InputSection.cpp, can you put the +4  in the addend?
>
> Where do you suggest to fix addend? It is impossible to do in the
> getImplicitAddend routine because we do not pass SymbolBody to this
> function. Without getImplicitAddend signature changing I can adjust
> addend in the place where we call getImplicitAddend (Writer.cpp).

At the call site in Writer.cpp should be fine.

Thanks,
Rafael