Re: [PATCH] also copy symbol attributes when equating an undefined symbol

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

On Wed, Nov 18, 2009 at 7:15 AM, Jan Beulich <JBeulich@novell.com> wrote:
>>>> "H.J. Lu" <hjl.tools@gmail.com> 18.11.09 16:01 >>>
>>On Wed, Nov 18, 2009 at 6:43 AM, Jan Beulich <JBeulich@novell.com> wrote:
>>>>>> "H.J. Lu" <hjl.tools@gmail.com> 18.11.09 15:32 >>>
>>>>On Wed, Nov 18, 2009 at 4:22 AM, Jan Beulich <JBeulich@novell.com> wrote:
>>>>> x86 has in md_estimate_size_before_relax() a check determining whether
>>>>> a symbol reference can go without relocation, which for 2.20 got added
>>>>> a check for GNU ifunc symbols. The linker tests would appear to
>>>>> exercise this code, but in reality the fact that ld-ifunc/lib.c gets
>>>>> translated correctly merely was due to __GI_library_func2 also being
>>>>> global. The reason is that this symbol is an equate of an (at the time
>>>>> of the equation at least) undefined symbol, and for such symbol
>>>>> attributes weren't copied so far in pseudo_set().
>>>>
>>>>Do you have a testcase which fails without your patch?
>>>
>>> No, I noticed the issue on the existing test case because of some small
>>> addition I'm using on top of the "official" code. I would have considered
>>> submitting that change eventually, but I hardly can if it breaks a test...
>>>
>>
>>So without your local change, you can't find any failing testcases?
>
> I just didn't have the time to try modifying the existing one so that it
> would expose the issue.
>

I don't think it is a goo idea to change assembler with a testcase:

1. You don't know for sure that it fixes the problem.
2. The problem may come back with other unrelated
changes.


-- 
H.J.