GH-128682: Stronger checking of PyStackRef_CLOSE and DEAD.

[markshannon]

Stronger checking of the stack state for PyStackRef_CLOSE. Only allow references to be closed from top to bottom, to ensure stack is valid when closing a reference.

Fix up bytecodes.c accordingly.

• Issue: Consistently spill the evaluation stack when making escaping calls #128682

[brandtbucher]

My impression was that, rather than enforcing a closing order, we would instead have the code generator insert NULL assignments to the stack where necessary. I vaguely remember situations where that would be really useful, such as when pushing or popping multiple stack items that are created or freed in the "wrong order", or where overlapping lifetimes would force us to convert things to heap references (and back again) to move things around on the stack.

Are we abandoning that idea, or is this a step towards that?

(Either way, the code in this PR seems correct.)

[brandtbucher]

What's the reasoning behind this alias? I see we do this in a couple of places now...

If the reasoning is that the analysis chokes on array accesses, maybe we should teach it to ignore them?

[markshannon]

We modify arguments, but the code generator assumes that args is a fixed array on the stack.

[brandtbucher]

Would it be useful to have this one "inherit" it's escaping-ness from its second arg? I could see us adding something for tuples or whatever and breaking this by accident.

[markshannon]

done

[markshannon]

My impression was that, rather than enforcing a closing order, we would instead have the code generator insert NULL...

That only applies when several references are removed from the stack at once, by DECREF_INPUTS.

If references are popped individually, they must be popped in the right order.