GH-93143: Don't turn LOAD_FAST into LOAD_FAST_CHECK

[brandtbucher]

The compiler's dataflow analysis for fast locals can be invalidated by jumps or deletions while tracing. When this happens, rather than changing co_code (which users assume to be constant) to use the less-efficient LOAD_FAST_CHECK everywhere, just display a RuntimeWarning and set any problematic locals to None instead.

• Issue: Dataflow analysis in the compiler to avoid runtime NULL checks #93143

[sweeneyde]

I can look closer this evening, but mutating co_code is not ideal, so I think this is probably a good idea.

[markshannon]

There are two cases when we need to set locals to None.

1. Deletion of locals
2. Jumping to arbitrary code

In the first case we can set the local to None and issue the warning cheaply and probably correctly. That's fine.

However when jumping, trying to decide which locals need to be set to None is going to be complex, expensive and error prone.
So I think we should go with @sweeneyde suggestion of setting all the undefined local variables to None.
Either don't issue any warning (as it might be confusing), or maybe issue a generic warning only if any locals were set to None.

[brandtbucher]

I've updated the code to skip scanning the bytecode for LOAD_FAST. Any unbound locals will be assigned None on jumps, and (if so) a single RuntimeWarning will be issued for all of them.

[markshannon]

👍
This feels safer and more robust than what we have now.

[markshannon]

There aren't necessarily incorrect, so maybe just "locals that are unbound after ..."

[markshannon]

Suggested change

	Py_XINCREF(value);
	Py_INCREF(value);

[bedevere-bot]

When you're done making the requested changes, leave the comment: I have made the requested changes; please review again.