It's important to realise that an LLVM backend is not a panacea -- the OCaml native code generator is highly optimised for the calling conventions and OCaml GC already. LLVM's IR doesn't quite track values at the same abstraction level as OCaml so some features such as exceptions become very expensive if mapped onto LLVM without changes to the IR.

Another major recent advance in OCaml 4.03 (released last month) is the flambda middle layer -- basically an epic inlining and allocation elision pass. See https://blogs.janestreet.com/flambda/ for more on it, but it's already giving 10-20% performance improvement on real world code.

Multicore is also coming soon -- see https://ocaml.io/w/Multicore for a summary of activity there.