GH-119866: Spill stack pointer when making "escaping" calls.

[markshannon]

This PR:

• Changes the stacktop int offset field in the _PyInterpreterFrame to a stackpointer pointer.
• Wraps (almost?) all escaping calls in the interpreter in a STORE_SP() ... LOAD_IP() pair to spill and reload the stack pointer.
• Adds an assert to verify that any call that reenters the interpreter has spilled the stack pointer.

• Issue: Spill the stack pointer across calls in the interpreter. #119866

[markshannon]

Benchmarking shows no change in performance.

[mdboom]

If only used in debug builds, should this be inside #ifdef Py_DEBUG?

[brandtbucher]

I think debug builds have the same ABI compatibility as release builds these days. I know we like to pretend that PyThreadState is opaque, but this seems like it might case some nasty bugs if we do this.

[bedevere-bot]

🤖 New build scheduled with the buildbot fleet by @iritkatriel for commit 2bd1339 🤖

If you want to schedule another build, you need to add the 🔨 test-with-refleak-buildbots label again.

[iritkatriel]

LGTM.

I'd wait for the refleaks tests (or run them locally if they don't kick off).

[dfremont]

Shouldn't this (and the corresponding line in the next function) be #ifdef Py_DEBUG? I don't think the assertion added to _Py_Dealloc will ever see an updated value of sp_cached otherwise.

[markshannon]

Thanks.

[markshannon]

Given that this will conflict horribly with #118450 and we will need to add tooling to check that all escaping calls are bounded by a save/load sp pair, I'm closing this for now.

[Fidget-Spinner]

@markshannon quick question: would it be possible to use the cases generator to automatically insert this? As part of free threaded specializing interpreter this also has a use case.

[markshannon]

would it be possible to use the cases generator to automatically insert this?

It might be possible, but I think it better to use the tools to verify that all escaping calls are wrapped.
Rather than SAVE_IP(); ...; LOAD_IP() we can add a macro ESCAPING_CALL(...) to make it both easier to write and to verify.

As part of free threaded specializing interpreter this also has a use case.

We can do this once #118450 is done.