gh-131798: constant fold classmethod and staticmethod in JIT

[kumaraditya303]

• Issue: Better uop coverage in the JIT optimizer #131798

Script used for benchmark:

"""Benchmark for classmethod/staticmethod JIT optimization."""

import time

LOOPS = 10_000_000


class MyClass:
    @classmethod
    def class_method(cls):
        return cls

    @staticmethod
    def static_method():
        return 42

    def regular_method(self):
        return self


obj = MyClass()


def bench_classmethod():
    o = obj
    for _ in range(LOOPS):
        o.class_method()
        o.class_method()
        o.class_method()
        o.class_method()
        o.class_method()


def bench_staticmethod():
    o = obj
    for _ in range(LOOPS):
        o.static_method()
        o.static_method()
        o.static_method()
        o.static_method()
        o.static_method()


benchmarks = [
    ("classmethod", bench_classmethod),
    ("staticmethod", bench_staticmethod),
]

for name, func in benchmarks:
    t0 = time.perf_counter()
    func()
    dt = time.perf_counter() - t0
    print(f"{name:20s} {dt:.3f}s")

Benchmark	main	PR	Speedup
classmethod	1.000s	0.572s	1.75x faster
staticmethod	0.860s	0.482s	1.78x faster

[kumaraditya303]

This can be done for class lookups as well so things like Foo.bar() where bar is a classmethod gets constant folded.

[Fidget-Spinner]

I don't see any change in deltablue, but that's because it's using cls.classmethod() instead of self.classmethod(). Would you like to add that to this PR as well? https://github.com/python/pyperformance/blob/main/pyperformance/data-files/benchmarks/bm_deltablue/run_benchmark.py#L234

[Fidget-Spinner]

I don't see any change in deltablue, but that's because it's using cls.classmethod() instead of self.classmethod(). Would you like to add that to this PR as well? https://github.com/python/pyperformance/blob/main/pyperformance/data-files/benchmarks/bm_deltablue/run_benchmark.py#L234

This turned out to be harder than I thought, because it would fold the classmethod descriptor into a unbound method lookup rather than a bound one, which changes the following CALL_BOUND_METHOD_EXACT_ARGS in the trace.

[NekoAsakura]

The outer else doesn't set self_or_null[0], intentional?

[markshannon]

Suggested change

	ADD_OP(_POP_TOP, 0, 0);
	optimize_pop_top(ctx, this_instr, owner);

[markshannon]

I don't see any change in deltablue, but that's because it's using cls.classmethod() instead of self.classmethod(). Would you like to add that to this PR as well? https://github.com/python/pyperformance/blob/main/pyperformance/data-files/benchmarks/bm_deltablue/run_benchmark.py#L234

This turned out to be harder than I thought, because it would fold the classmethod descriptor into a unbound method lookup rather than a bound one, which changes the following CALL_BOUND_METHOD_EXACT_ARGS in the trace.

That shouldn't be a problem for staticmethod though, as self_or_null is always NULL.

It would make sense for LOAD_ATTR 1 for classmethod to be leaving func cls on the stack.

If we special case classmethod in _PyObject_GetMethodStackRef, it should add negligible overhead in the interpreter (possibly speeding things up if classmethods are common) and will enable this optimization in the JIT.

[Fidget-Spinner]

If we special case classmethod in _PyObject_GetMethodStackRef, it should add negligible overhead in the interpreter (possibly speeding things up if classmethods are common) and will enable this optimization in the JIT.

Working on that.