[llvm-dev] Permitted success/failure orderings for atomic compare_exchange

[Oliver Giersch via llvm-dev]

A question has come up on how to interpret the wording of LLVM's documentation regarding the possible memory ordering for success and failure of atomic compare_exchange operations.

From the LLVM reference:

"The success and failure ordering arguments specify how this cmpxchg synchronizes with other atomic operations. Both ordering parameters must be at least monotonic , the ordering constraint on failure must be no stronger than that on success, and the failure ordering cannot be either release or acq_rel ."

The way this operation is implemented in  the standard library of the Rust programming language, the combination of release (success) and acquire (failure) is not permitted and will fail at runtime, the prevailing opinion being, that a failing CAS with release/acquire ordering is actually relaxed on the store part and hence acquire ordering for the failure case would be stronger.

However, I don't believe that this a correct reading of the specification and that neither acquire nor release should be considered stronger or weaker than the other and that hence that combination should be permitted.

I had hoped to get get some clarification on this issue here.

[Eli Friedman via llvm-dev]

The LLVM source interprets it as https://github.com/llvm/llvm-project/blob/a5d375e0cbc4abf6e318270a1bc9c1c0961ae565/llvm/include/llvm/Support/AtomicOrdering.h#L89 .

 

We should probably drop the requirement; the C++ standard dropped the corresponding requirement for the library call.  See http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2016/p0418r2.html .

 

-Eli

[Oliver Giersch via llvm-dev]

Thanks for the hints, the text of the standards proposal was very insightful. I am not entirely sure how to interpret the linked source code. So does the current (previous?) or rather LLVM's current interpretation/implementation of that standard consider the combination release on success and acquire on failure to be legal? The table lists both relationships "release isStrongerThan acquire" and "acquire isStrongerThan release" as false. However, the table below it shows the "isEqualOrStrongerThan" relationship to also be false.

[Oliver Giersch via llvm-dev]