The first-working version of parallelism in HLVM was very inefficient due to repeated fetching of thread-local data via pthreads. The latest update passes thread-local data to every function in an auxiliary argument and achieves production-level performance: 2× faster than OCaml on average over non-list benchmarks.