Additional race introduced in variable microbench compared to original about dataracebench HOT 7 CLOSED

The race on j introduces another spurious write-write race in the next line:
With the race on j the loop invariant j<len is no longer valid, so that we see write-write races between a[i][len] and a[i+1][0]. Although this is an interesting race to detect, I think this race deserves an own kernel and documentation.
For i=len-2, this can also lead to out-of-bound read as in a[len-1][len].

from dataracebench.

Same problem happens in the antidep2-orig-yes.c benchmark.

from dataracebench.

There is a similar problem also in the benchmark plusplus-orig-yes.c.

Because of the race in the index variable "outLen", it may introduce a race also on the array output.

from dataracebench.

And truedepfirstdimension-orig-yes.c

from dataracebench.

@simoatze for plusplus-orig-yes.c I think, this is the intended behavior of the kernel, although this side effect is not documented in the description.

I created a PR to solve above issues

from dataracebench.

@jprotze and @simoatze Thanks a lot for filing this issue. We will incorporate your suggested fixes into our next release (along with a few additional benchmark programs) soon.

from dataracebench.

This issue has been resolved as part of our recent release v1.1.0. Thanks for reporting this to us ( and with fixes)!

from dataracebench.