Calling methods and lookup up attributes when receiver is `super()` has extra cost comparing to regular attribute lookup. It mainly comes from the need to allocate and initialize the instance of the `super` which for zero argument case also include peeking into frame/code object for the `__class__` cell and first argument. In addition because `PySuper_Type` has custom implementation of tp_getattro - `_PyObject_GetMethod` would always return bound method. 
```
import timeit

setup = """
class A:
    def f(self): pass
class B(A):
    def f(self): super().f()
    def g(self): A.f(self)
b = B()
"""
print(timeit.timeit("b.f()", setup=setup, number=20000000))
print(timeit.timeit("b.g()", setup=setup, number=20000000))

7.329449548968114
3.892987059080042
```
 
One option to improve it could be to make compiler/interpreter aware of super calls so they can be treated specially. Attached patch introduces two new opcodes LOAD_METHOD_SUPER and LOAD_ATTR_SUPER that are intended to be counterparts for LOAD_METHOD and LOAD_ATTR for cases when receiver is super with either zero or two arguments.

Immediate argument for both LOAD_METHOD_SUPER and LOAD_ATTR_SUPER is a pair that consist of:
 0: index of method/attribute in co_names
 1: Py_True if super was originally called with 0 arguments and Py_False otherwise.

Both LOAD_METHOD_SUPER and LOAD_ATTR_SUPER expect 3 elements on the stack:
TOS3: global_super
TOS2: type
TOS1: self/cls

Result of LOAD_METHOD_SUPER is the same as LOAD_METHOD.
Result of LOAD_ATTR_SUPER is the same as LOAD_ATTR

In runtime both LOAD_METHOD_SUPER and LOAD_ATTR_SUPER will check if `global_super` is `PySuper_Type` to handle situations when `super` is patched. If `global_super` is `PySuper_Type` then it can use dedicated routine to perform the lookup for provided `__class__` and `cls/self` without allocating new `super` instance. If `global_super` is different from `PySuper_Type` then runtime will fallback to the original logic using `global_super` and original number of arguments that was captured in immediate.

Benchmark results with patch:
4.381768501014449
3.9492998640052974

Why?

Do you have any evidence that the overhead of super() is significant in real programs, or that the proposed change actually speeds up anything beyond your micro-benchmark?

Currently, super() is decoupled from the core language.  It is just a builtin that provides customized attribute lookup.  This PR makes super() more tightly integrated with the core language, treating it as if it were a keyword and part of the grammar.  Also note, users can currently create their own versions of super(), shadowing the builtin super().

>Currently, super() is decoupled from the core language.  It is just a builtin that provides customized attribute lookup.  This PR makes super() more tightly integrated with the core language, treating it as if it were a keyword and part of the grammar.  Also note, users can currently create their own versions of super(), shadowing the builtin super().

This is true however:
- this patch does not block people from introducing custom version of `super` so this scenario still work. The idea was to streamline the common case
- based on digging into Instagram codebase and its transitive dependencies (which is reasonably large amount of code) all spots where `super()` appear in sources assume `super` to be builtin and for a pretty common use-case its cost is noticeable in profiler.
- zero-argument `super()` still a bit magical since it requires compiler support to create cell for `__class__` and assumes certain shape of the frame object so this patch is a step forward with a better compiler support and removing runtime dependency on the frame

> Do you have any evidence that the overhead of super() is significant in real programs

I do see the non-negligible cost of allocation/initialization of `super` object in IG profiling data.

This looks like a sensible idea to me. The safeguards to ensure that customized 'super' still works seem reasonable to me. I have to admit that I sometimes refrain from using super() where I should because of the expense, so this would be welcome.

I do wonder -- is two-arg super() important enough to support it at all? Maybe the code would be somewhat simpler if the special opcodes were only generated for zero-arg super() calls.

Numbers please.

What is "non-negligible cost of allocation/initialization" mean as a fraction of runtime?
What sort of speed up are you seeing on whole programs?

Apologies for the delay in reply: in more concrete numbers for IG codebase enabling this optimization resulted in 0.2% CPU win.