As I said, it needs experimentation. The performance characteristics of the interpreter are very different from compiled code so I kind of doubt that our MIR optimizations will make a significant difference. But I might be completely wrong about this.
Basic inlining of small functions, constant folding, and dead code elimination could help immensely in some cases (e.g. in indexing-heavy code). I had a case where I had to write a specialized version of a function for const eval, just because the same computation without manual optimizations was exceeding the const eval limit.