(co-author of the OCaml memory model paper here)

The details of the 'LDRF' (local data race freedom) property are described in detail here: https://anil.recoil.org/papers/2018-pldi-memorymodel.pdf

The performance numbers are in the paper abstract: "our evaluation demonstrates that it is possible to balance a comprehensible memory model with a reasonable (no overhead on x86, ~0.6% on ARM) sequential performance trade-off in a mainstream programming language". It's a little higher on PowerPC but still very usable, and RISC-V overheads should be roughly comparable to ARM.

As the paper clarifies, this is for the performance of non-atomic read/writes in sequential setting. The paper left the performance evaluation for atomic read/writes to future work. Is there any indication yet regarding the performance in a parallel setting compared to weaker guarantees?

That's right -- we first wanted to establish that existing OCaml code wouldn't be adversely impacted. There are various efforts ongoing to build interested (locked and lock-free) concurrent data structures, so that will inform the parallel performance results. Nothing published yet.

Does this feature inhibit any kind of optimization that current compilers could perform?

According to the paper “any compiler optimisation that breaks the load-to-store ordering is disallowed.”