Disabling cilksan on a specific line

[wheatman]

I have some code in which I use some per worker data by creating a vector and indexing in to it with __cilkrts_get_worker_number() which I know is not the best practice, but I have found it a useful one in some cases.

Is there a way I can tell cilksan to not worry about a specific line of code, or a specific memory address or variable so that I don't get a bunch of warnings about these lines which are not what I am looking for when I am looking for bugs.

[neboat]

Cilksan has an API you can use by including the cilk/cilksan.h header into your code. Depending on what exactly you want to do, the API provides a couple of options.

Cilksan fake locks

Cilksan includes an API for fake locks, which direct Cilksan to treat memory accesses as if they are protected by mutex lock. Cilksan fake locks are fake, in that they do not actually protect memory accesses with a mutex lock or provide any atomicity guarantee. Instead, they merely direct Cilksan to treat code as if it is protected by a lock. Cilksan will therefore not report races on code if logically parallel accesses are protected by the same fake lock.

The API for Cilksan fake locks is based on that for ordinary mutex locks. Here's a simple C++ example:

double wls_sum[__cilkrts_get_nworkers()];
Cilksan_fake_mutex fake_lock;
cilk_for (int i = 0; i < n; ++i) {
  __cilksan_acquire_lock(&fake_lock);
  wls_sum[__cilkrts_get_worker_number()] += A[i];
  __cilksan_release_lock(&fake_lock);
}

The API also provides Cilksan_fake_lock_guard for C++ code that you can use to automatically acquire and release a fake lock for a lexical scope, analogously to std::lock_guard. For example:

double wls_sum[__cilkrts_get_nworkers()];
Cilksan_fake_mutex fake_lock;
cilk_for (int i = 0; i < n; ++i) {
  Cilksan_fake_lock_guard guad(&fake_lock);
  wls_sum[__cilkrts_get_worker_number()] += A[i];
}

Cilksan fake locks have zero cost when your code is not compiled with Cilksan.

Cilksan disable/enable checking

The Cilksan API also provides the functions __cilksan_disable_checking() and __cilksan_enable_checking() to simply disable or enable Cilksan for a region of code. Note that these functions will cause Cilksan to completely ignore part of the code, thus providing no race-detection on those operations whatsoever. In contrast, when Cilksan fake locks are used, Cilksan will still record memory accesses that are protected by a fake lock, but it won't report races when parallel instructions are protected by a common fake lock. Therefore, if your program performs some unintended parallel access to that memory, that parallel access will likely not be protected by the same fake lock, and Cilksan will still report a race, helping you find that bug. That's why I recommend using Cilksan fake locks instead of the routines to disable and enable checking.

Hope that helps. Please let us know if you have more questions or run into issues.

[wheatman]

This worked perfectly,
Thanks.
Is there any spot to check for documentation for the cilk tools in general?