Rationalize and complete the main JIT optimizer pass

[markshannon]

The main JIT optimizer pass is a bit untidy at the moment. It lacks coherent design and is split into two parts, which makes it hard to work with.

We should enhance the coherency to simplify future maintenance and ease in new contributors.
To do this, we need to do the following:

• Use symbols consistently throughout, instead of the current mixture of symbols , ad-hoc variables and references to objects.
• Move the "globals to constants" pass into the main optimizer pass
• Move the state maintained by local variables in the optimizer into structs within the optimizer context for clarity and to avoid surprising action at a distance effects.

Linked PRs

• GH-138378: Move globals-to-consts pass into main optimizer pass #138379

[markshannon]

@Fidget-Spinner
@brandtbucher

[brandtbucher]

Agreed. Would “get rid of stack allocated arenas” and “don’t operate on the uop stream in-place” belong here too?

[Fidget-Spinner]

“don’t operate on the uop stream in-place” belong here too?

This is surprisingly easy to do. The WIP partial eval pass on my end does that and it isn't as annoying as I expected.

[Fidget-Spinner]

We also have one open question regarding rationalizing the pass:

If we bail early, should we just end the trace there (override the current instruction with _EXIT_TRACE ?). There's quite a few opcodes where we end early (such as generators, complex functions, or stuff we haven't supported yet in the optimizer). The current behaviour is to just let the whole unoptimized trace get compiled. However, I feel that this is an oversight, as I recall the original plan being anything we can't optimize, we just bail to the interpreter? Should I change that?