Tier 2 abstract interpreter for the optimizer

[Fidget-Spinner]

Feature or enhancement

(Mega issue) Start doing optimization passes on tier 2 bytecode.

Pitch

The target of tier 2 optimizations is tier 2 bytecode. Abstract interpretation is a natural way to analyze and optimize tier 2 bytecode. We can generate the abstract interpreter from the bytecode DSL

Previous discussion

See faster-cpython/ideas#611

Todo list:

• Generate a barebones tier 2 abstract interpreter from the DSL.
• Set up the abstract interpreter. This requires that we pass in the runtime state at the point of entering the executor.
• Perform partial evaluation. For now, all static comes from LOAD_CONST. Assume everything else is dynamic.

The initial partial evaluation will be bad because it does not have that much static information (we need watchers for methods, functions, global, etc., to make them effectively static). However, that phase can be done in the region formation step before partial evaluation. Someone else can pick them up as a parallel workstream.

Linked PRs

• gh-107557: Tier 2 barebones abstract interpreter #107558
• gh-107557: Setup abstract interpretation #107847
• gh-107557: Remove unnecessary SAVE_IP instructions #108583

[Fidget-Spinner]

I now have a clearer idea of what I'm going to do. I'm abandoning this issue and re-creating a new one since most of the description is outdated.