But we are already able to extract parallelism deep down in the instruction level with out of order execution.

Imagine some layer in between an abstract syntax tree and micro-ops. Again, just thinking out loud.