@markshannon this needs benchmarking

Does this really fix the unsafe decrefs, or just some of them?

It isn't just the _Py_DECREF_SPECIALIZED, it is also the DECREF_INPUTS_AND_REUSE_FLOAT and _Py_DECREF_NO_DEALLOC. (I see you've also removed the _Py_DECREF_NO_DEALLOC)

I can also see one unsafe Py_DECREFs in RERAISE, there may be more.

this needs benchmarking

There's not much point if we already know it's going to be slower.
We should replace _Py_DECREF_SPECIALIZED with PyStackRef_CLOSE_SPECIALIZED to avoid the slowdown.

I can also see one unsafe Py_DECREFs in RERAISE, there may be more.

I manually went through every Py_DECREF in bytecodes.c. That Py_DECREF is safe. It's a PyStackRef_AsPyObjectSteal then Py_DECREF, not a borrow.

it is also the DECREF_INPUTS_AND_REUSE_FLOAT ~and

I don't know how to fix that without tanking performance. The way they behave doesn't conform to the stackref semantics at all.

Failures are the usual suspects.

Performance is a wash, as I would have expected.

So, lets' leave the DECREF_INPUTS_AND_REUSE_FLOAT macros for another PR and merge this.