Re: [PATCH][BZ 18743] PowerPC: Fix a race condition when eliding a lock

["Tulio Magno Quites Machado Filho" <tuliom at linux dot vnet dot ibm dot com>]

Peter Bergner <bergner@vnet.ibm.com> writes:

> On Thu, 2015-07-30 at 23:06 -0300, Tulio Magno Quites Machado Filho wrote:
>> Adhemerval Zanella <adhemerval.zanella@linaro.org> writes:
>> >> +    else								\
>> >> +      for (int i = __elision_aconf.try_tbegin; i > 0; i--)		\
>> >> +	{								\
>> >> +	  asm volatile("" ::: "memory");				\
>> >
>> > I can't really understand the requirement of this compiler barrier here.  If
>> > compiler is moving around the 'is_lock_free' *before* __builtin_tbegin IMHO
>> > this is a compiler bug.
>> 
>> This is the second problem and I agree with you again.  IMHO,  __builtin_tbegin
>> starts a critical path and the compiler should not be moving a memory access out
>> of the critical path or into the critical path.
>> However, as the current implementations of GCC have this "issue" (some people
>> may not agree with us), I believe we could carry this compiler barrier at least
>> until all GCC versions supporting __builtin_tbegin are fixed.
>
> I completely agree we need a barrier here, so I will fix that.  The question
> I'm starting to wonder, is do we need more than just a memory barrier or
> do we need a complete optimization barrier, so no code (even scalar code)
> will move past the __builtin_tbegin() (and other HTM builtins)?
> ...
> observable if one or more of the transactions succeeds.  Torvald,
> Adhemerval and Tulio, do you guys agree with that assessment?

I agree with you.
The compiler shouldn't move instructions that affect memory or any of the
checkpointed registers.

-- 
Tulio Magno