Add effective peephole optimizer for uop traces.

[markshannon]

Feature or enhancement

Currently we lack an effective peephole optimizer for uops traces.
We do some ad-hoc removal of LOAD/POP pairs, but many enhancements are missed.

For example, this silly program:

def add(a, b):
    return a + b

def loop(n=100_000):
    t = 0
    for _ in range(n):
        t += add(1,2)

if __name__ == "__main__":
    loop()

generates this trace:

which contains the sequence:

_LOAD_FAST_BORROW 0
_LOAD_FAST_BORROW_1
_LOAD_CONST_INLINE_BORROW K 
_RROT_3
_POP_TOP_NOP
_POP_TOP_NOP

which should be reduced to

_LOAD_CONST_INLINE_BORROW K 

Linked PRs

• gh-152215: Add uop peepholes for folded stack shuffles and adjacent guards #153035