Re: [PATCH] Fix sporadic failure in g++.dg/tsan/aligned_vs_unaligned_race.C

[Mike Stump <mikestump@comcast.net>]

On Jan 19, 2015, at 12:47 AM, Dmitry Vyukov <dvyukov@google.com> wrote:
> Long story short. Tsan has a logical data race the core of data race
> detection algorithm. The race is not a bug, but a deliberate design
> decision that makes tsan considerably faster.

Could you please quantify that for us?  Also, what lockless update method did you use?  Did you try atomic increment?  On my machine, they are as cheap as stores; can’t imagine they could be slow at all.  If the latency and bandwidth of atomic increment is identical to store, would the costs be any higher than using a store to update the tsan data?  A proper port of tsan to my machine would make use of atomic increment.  I consider it a simple matter to sequence the thread termination and the output routine to ensure that all the updates in the threads happen before the output routine runs.  The output routine strikes me as slow, and thread termination strikes me as slow, so taking a little extra time there seems reasonable.  Was the excessive cost you saw due to the termination costs?

> So ironically, if the race memory accesses happen almost simultaneously, tsan can miss the
> race.
> Thus we have sleeps.

I’ve not seen a reason why the test suite should randomly fail.  The gcc test suite does not.  Could you explain why the llvm test suite does?  Surely you know that sleep is not a synchronization primitive?

> Sleeps vs barrier is being discussed in the "Fix parameters of
> __tsan_vptr_update" thread.

When finished, let us know the outcome.  To date, I’ve not seen any compelling reason to have the core of tsan be other than deterministic and the test suite other than deterministic.  I’d love to see the backing for such a decision.

> I would really like to keep llvm and gcc tests in sync as much as possible.

Excellent, from my perspective, that would mean that you make the llvm test suite deterministic.