This comes from #2622 (comment)

Why mp_call_function_n_kw() wasn't originally implemented in such a way as to require a slot for self at the beginning of args, then skip it if not needed? Unless there were good reasons, let's refactor that now ;-).

Please elaborate on this idea. Do you mean that mp_call_function_n_kw() should take an extra argument which is the first argument, like mp_call_function_n_kw(fun, arg0, n_args, n_kw, args). Or that the array args should contain an empty slot at args[0]? If the latter then, well, ignoring the fact that this array is const (and I think it's const for a good reason), mp_obj_instance_call won't be able to use the empty slot because the same incoming array is then passed to mp_call_function_n_kw, which requires another empty slot.

Or that the array args should contain an empty slot at args[0]?

I meant this.

mp_obj_instance_call won't be able to use the empty slot because the same incoming array is then passed to mp_call_function_n_kw, which requires another empty slot.

So, per the idea above, it's not "empty" slot, but "self" slot. There can be only 1 self argument, so it's not recursive, but static. I looked at what it would take to add it, and changes are indeed substantial. But what I somehow assumed that a lot of code paths would be affected by this issue, but looking at the code reminded that everything what's possible was already optimized, specifically expressions of the form obj.symbol(...). So, the problem (specifically, a need to semantically "explode" obj(...) into obj.prop(obj, ...)) affects only 2 cases: meth = obj.meth; ... ; meth(...) and obj(...). Expressed like that, it was easy to find a good solution - to reuse the same code paths for both cases: 037e691