STORE_FAST__LOAD_FAST_SAME

[sweeneyde]

I mentioned this here, but here's a possible super-instruction for when STORE_FAST is followed by a LOAD_FAST with the same oparg.

        TARGET(STORE_FAST__LOAD_FAST_SAME) {
            PyObject *value = TOP();
            Py_INCREF(value);
            SETLOCAL(oparg, value);
            next_instr++;
            NOTRACE_DISPATCH();
        }

Benefits:

• no need to decode second oparg
• elide the pop-then-push stack manipulations
• no need to check for null in LOAD_FAST

I haven't benchmarked yet, but I have collected stats for python -m test and stats for pyperformance.
The test suite resulted in:

Name	Count	Self
STORE_FAST__LOAD_FAST_SAME	90557905	1.6%
STORE_FAST__LOAD_FAST	87169944	1.6%

And pyperformance resulted in

Name	Count	Self
STORE_FAST__LOAD_FAST	618849351	2.6%
STORE_FAST__LOAD_FAST_SAME	413324475	1.7%

This might also be beneficial in some comprehensions.

[brandtbucher]

Heh, I had the same idea and tried this a few weeks ago:

python/cpython@main...brandtbucher:lf-lf

I wasn't able to move the benchmarks, though (even when adding an additional LOAD_FAST__LOAD_FAST_SAME opcode on top of it).