Further improve performance

[tim-hoffman]

No description provided.

[0xddom]

I don't really understand what's going on, but I trust you that this improves performance. Because, if I understood it correctly, by splitting between compute and execute you avoid copying the env unnecessarily, correct? Do you have some numbers of the run time now?

Also, I was thinking. Do we still clone the whole IR every time we run a pass? Maybe we could improve performance further if we move the IR and only create new copies of things if we actually modify them. I don't know if it will work since we need the IDs for each bucket but may be worth a try

[0xddom]

nit: typo

[tim-hoffman]

You got it. Profiling showed that Env clones and destructors were taking the most time by far.
I've replaced "execute" with "compute" in many cases where the bucket cannot update the Env. The compute_or_execute macro helps in cases where the Env might be updated but we don't end up using the updated Env so it's safe to try the "compute" approach first but if an Env modifying instruction is encountered, fall back to the "execute" approach.

[tim-hoffman]

Each pass does still reconstruct the IR via the transform methods but I haven't seen that as a bottleneck in profiling so far.

[0xddom]

Each pass does still reconstruct the IR via the transform methods but I haven't seen that as a bottleneck in profiling so far.

It's not only performance but memory usage, but we can leave it for future work then, if it's not as pressing. This PR has enough stuff already