[PATCH] nptl: Fix invalid Systemtap probe in pthread_join [BZ #24211]

[Carlos O'Donell]

On 2/11/19 4:40 PM, Florian Weimer wrote:
> * Carlos O'Donell:
> 
>>> -  LIBC_PROBE (pthread_join_ret, 3, threadid, result, pd->result);
>>> +  LIBC_PROBE (pthread_join_ret, 3, threadid, result, pd_result);
>>
>> What prevents the compiler from optimizing away pd_result and using
>> pd->result directly again?
> 
> The __free_tcb call acts as an implicit optimization barrier, like any
> function call that doesn't have attributes that say otherwise.

Sorry, yes, that's right, global memory accesses should not move around
that call, so I guess that means that the original pd_result store wont
move past __free_tcb() (and we don't do IPO/LTO yet).

>> The actions of __free_tcb() and their consequences are not visible
>> to the compiler.
> 
> Exactly.  The compiler cannot rematerialize the value from pd->result
> because it has no way of knowing if the value is still the same.

Yes, OK, let me review this again.

-- 
Cheers,
Carlos.