Profiling tools are post-mortem, and often not able to be enabled in production. What about when the code is not yet running, and you're trying to reason about

Profiling tools are post-mortem, and often not able to be enabled in production. What about when the code is not yet running, and you're trying to reason about it as you write it? What about when the performance problem is workload dependent, and only emerges in production, and you need to reason inductively from the source code about what could be going wrong?