[llvm] [RFC][LangRef] Specify that the accessed bytes of concurrent atomics must be either disjoint or the same (PR #204329)

[Fabian Ritter via llvm-commits]

ritter-x2a wrote:

> But ISTM that it might also still be useful to clarify the wording in "Atomic Memory Ordering Constraints" to indicate that none of these effects occur unless the memory access size is consistent between the operations. While it may already be implied by a close parsing of the newly-added bullet point here, writing that more clearly would probably be helpful.

@jyknight I've extended the PR to clarify that the "Atomic Memory Ordering Constraints" (and fence synchronization, which is specified separately) only apply to accesses that perfectly overlap.

(Also: feel free to bike-shed on the term "perfectly overlapping".)

https://github.com/llvm/llvm-project/pull/204329