On Sun, 4 Jan 2015 13:57:38, Mike Stump wrote:
>
> On Jan 4, 2015, at 1:48 PM, Bernd Edlinger <bernd.edlin...@hotmail.de> wrote:
>> I would need a way to link the test case two helper functions, that are not
>> compiled with -fsanitize=thread
>
> /* { dg-additional-sources “tsan-helper.c" } */
>
> might be one step forward to do that. I don’t know off hand about the
> options. I think you may be able to specify unique options per file.
>
>> I tried, it and it works 10.000 times without one failure.
>
> So, statistically knowing it works in practice is nice. However, the standard
> is, does is work by design? I’ll let you comment on that. The sync approach I
> designed to work in any case, no matter the complexity, by design. I’d assume
> that you’ve engineered it to work by design.
IMO Yes, but Dmitry may know better. Tsan does not acquire a mutex _before_ it
detects a race.
And race conditions that happen exactly synchronous may be invisible.
If we use some hidden mechanism to ensure that certain events happen in a
certain sequence,
that would make a big difference. It would work much better than sleep(), we
only need sleep if
we need to see the "as if synchronized via sleep" diagnostic.
I see now, that our tsan tests are almost verbatim copies of the LLVM test
suite.
They _do_ have problems with the stability of positive tests, especially under
stress,
I see all positive tests are called by a "deflake" tool, that exercises the
test 10 times,
until it eventually succeeds.
I see one other test (which we apparently do not have here),
was changed recently from sleep(2) to sleep(4)
[Tsan] Fix the atomic_race.cc test to pass on systems with high loads
Differential Revision: http://reviews.llvm.org/D6478
--- compiler-rt/trunk/test/tsan/atomic_race.cc 2014/05/30 14:08:51 209898
+++ compiler-rt/trunk/test/tsan/atomic_race.cc 2014/12/02 15:04:39 223122
@@ -36,7 +36,7 @@
for (int i = 0; i < kTestCount; i++) {
Test(i, &atomics[i], false);
}
- sleep(2);
+ sleep(4);
for (int i = 0; i < kTestCount; i++) {
fprintf(stderr, "Test %d reverse\n", i);
Test(i, &atomics[kTestCount + i], false);
Thanks
Bernd.
