JIT: _INIT_CALL_BOUND_METHOD loses too much information/_RECORD_BOUND_METHOD not implemented

[Fidget-Spinner]

Feature or enhancement

Proposal:

It seems _RECORD_BOUND_METHOD exists in bytecodes.c but is not implemented in optimizer_bytecodes.c This means _INIT_CALL_BOUND_METHOD fails to get probable method and does not continue optimizing the trace.

The problem with not doing this is that the _PUSH_FRAME at the call site for the bound method decides to hit bottom because it sees no code object :(. So any trace with a call to a bound method does not get optimized after the call. This is pretty bad...

Has this already been discussed elsewhere?

No response given

Links to previous discussion of this feature:

No response

Linked PRs

• gh-148438: implement _RECORD_BOUND_METHOD in JIT #148457

[Fidget-Spinner]

@NekoAsakura would you like to take this on please (also, I know I've been giving you a lot of issues recently, but it's only because you're good :).

The idea is to change _RECORD_BOUND_METHOD to record the actual bound callable:

        tier2 op(_RECORD_BOUND_METHOD, (callable, self, args[oparg] -- callable, self, args[oparg])) {
            PyObject *callable_o = PyStackRef_AsPyObjectBorrow(callable);
            if (Py_TYPE(callable_o) == &PyMethod_Type) {
                RECORD_VALUE(callable_o);
            }
        }

In _CHECK_METHOD_VERSION, we promote the value to a constant one if func_version matches (we need to leave the guard in though). We promote it using sym_set_const after getting the sym_get_probable_value.

Finally, _EXPAND_METHOD will expand the method same as what it does in bytecodes.c, and sym_new_const the callable and self.

There should be a test that tests if we manage to optimize through a call to a bound method (e.g. A.b()). The trace inside the b should be optimized, and the optimizer should not stop midway.

[NekoAsakura]

NP! Keep them coming. Having a clear direction is much better than fumbling about, and I'm properly enjoying getting to know JIT through these PRs.