gh-120619: Tier 2 partial evaluator foundations

[Fidget-Spinner]

This PR sets up a tier 2 partial evaluation pass' foundations. It does the following:

1. Introduce binding-time analysis of instructions and locals.
2. Virtualized stack values (This was why we need a local slot now to indicate if the stack value is pointing to a virtualized instruction).
3. A new abstract interpreter -- one for partial evaluation.
4. Symbolic stack reification -- reconstructing the stack when we need to/hit a non-static instruction. This is done by telling the originating instruction to devirtualize itself.
5. It removes most pure annotations -- we need to reimplement those in the partial evaluator for every one we introduce. Otherwise it will do arbitrary optimizations on them.

As a litmus test, it does simple dead store elimination by tracking locals. For example, the code x = x is now optimized to a nop.

This is different than the previous PR at #123652. The main difference is that we create the residual in-place. By "looking backwards" at the originating instruction and marking that as virtual/non-virtual. If you want a visualization of how this will work. See Mark's talk at EuroPython 2023 here https://youtu.be/dgrtgtT-UXM?t=1198.

Also up for discussion: the generator for the partial evaluator is exactly the same as the specializer. Should I just remove it?

• Issue: Add tier 2 partial evaluation pass. #120619

[Eclips4]

Great work, Ken!

[markshannon]

I don't understand why any of bytecodes.c, optimizer_analysis.c, optimizer_bytecodes.c, optimizer_symbols.c, or the structs in pycore_optimizer.h need to change.
This is adding a new pass. There should ideally be no changes to the existing pass. If changes are necessary they should be few and small.

What does _static mean? Using "static" is very confusing unless it relates to the C "static" which describes scope.

Rather than removing the pure annotations, we should make sure that they are in the right places.
e.g..
LOAD_FAST is marked as pure, but it clearly has side effects.
BINARY_SUBSCR_STR_INT and BINARY_SUBSCR_TUPLE_INT are not marked as pure, but they are.

[Fidget-Spinner]

I don't understand why any of bytecodes.c, optimizer_analysis.c, optimizer_bytecodes.c, optimizer_symbols.c, or the structs in pycore_optimizer.h need to change.

This is adding a new pass. There should ideally be no changes to the existing pass. If changes are necessary they should be few and small.

I reused the infra for the old pass to build the new one. Though I can separate them totally for separation of concerns. Let me know what you think.

What does _static mean? Using "static" is very confusing unless it relates to the C "static" which describes scope.

It means potentially static in the context of binding-time analysis. It's more liberal than pure.

[Fidget-Spinner]

@markshannon I split it out into its own files, and own symbols, so it doesn't share the type information too. This simplifies the lattice quite a bit.