GH-102300: Reuse objects with refcount == 1 in float specialized binary ops.

[markshannon]

About 0.5% faster:

https://github.com/faster-cpython/benchmarking/tree/main/results/bm-20230227-3.12.0a5%2B-c29d369

The results are quite noisy.
But the results seem good, as the most float heavy benchmark, nbody, does show a 7% speedup.

• Issue: Reuse objects with refcount == 1 in specialized instructions #102300

[Fidget-Spinner]

I can't see the results. I think external parties can't view that repo.

[brandtbucher]

I can't see the results. I think external parties can't view that repo.

Sorry, you can change the repo name from benchmarking to benchmarking-public in the URL:

https://github.com/faster-cpython/benchmarking-public/tree/main/results/bm-20230227-3.12.0a5%2B-c29d369

[brandtbucher]

Looks good (like the clever use of _Py_DECREF_NO_DEALLOC). Do we want a helper function or macro to reduce duplication?

In the original approach (#30594), I only considered the LHS, and added a special case for in-place ops followed by some variant of LOAD_FAST (where the refcount is 2). Not sure if we want to consider adding something similar later.

[markshannon]

An inline function is tricky without introducing extra branches, so I've added a macro.

The refcount == 2 variant probably wants support from the bytecode compiler to mark candidates operations in x = x + (complicated expression).