Optimize reference counting overhead of `LOAD_FAST` variants

[mpage]

Feature or enhancement

Proposal:

LOAD_FAST and its super instruction form are the most frequently executed bytecode instructions in most Python programs (they represent ~20% of dynamic instruction frequency on the benchmark suite). While they are very cheap (they're just an incref and a push to the stack), they are not free. If we can prove that the reference in the frame outlives the reference that is loaded on the stack we can use a cheaper variant of LOAD_FAST that loads an appropriately tagged borrowed reference onto the operand stack, thereby avoiding the incref/decref.

Has this already been discussed elsewhere?

This is a minor feature, which does not need previous discussion elsewhere

Links to previous discussion of this feature:

faster-cpython/ideas#700

Linked PRs

• gh-130704: Strength reduce LOAD_FAST{_LOAD_FAST} #130708