Looks like the sqlite bench complaint was fixed 5 days ago and the react-hello benchmark was fixed earlier today (before the creation time of this gist).
I don't know if the suggested FFI benchmark is fair, as the change would no longer be simulating doing many ffi_hashes on many byteArrays, instead it simulates doing many ffi_hashes on the same byteArray over and over and benefits from caching the type casting necessary by deno.
Replace the const byte array with a randomly generated list (or real bytes you'd want to hash) of many passed to the benchmark as a parameter and the benchmark and the deno one loses the advantage from the cached type cast.
I don't know if the suggested FFI benchmark is fair, as the change would no longer be simulating doing many ffi_hashes on many byteArrays, instead it simulates doing many ffi_hashes on the same byteArray over and over and benefits from caching the type casting necessary by deno.
Replace the const byte array with a randomly generated list (or real bytes you'd want to hash) of many passed to the benchmark as a parameter and the benchmark and the deno one loses the advantage from the cached type cast.