Use dedicated opcodes to speed up calls/attribute lookups with super() as receiver

[vladima]

BPO	43563
Nosy	@gvanrossum, @rhettinger, @markshannon, @vladima, @isidentical
PRs	bpo-43563 : Introduce dedicated opcodes for super calls #24936

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = None
closed_at = None
created_at = <Date 2021-03-19.20:57:57.938>
labels = ['interpreter-core', '3.10']
title = 'Use dedicated opcodes to speed up calls/attribute lookups with super() as receiver'
updated_at = <Date 2021-04-27.18:23:03.781>
user = 'https://github.com/vladima'

bugs.python.org fields:

activity = <Date 2021-04-27.18:23:03.781>
actor = 'v2m'
assignee = 'none'
closed = False
closed_date = None
closer = None
components = ['Interpreter Core']
creation = <Date 2021-03-19.20:57:57.938>
creator = 'v2m'
dependencies = []
files = []
hgrepos = []
issue_num = 43563
keywords = ['patch']
message_count = 7.0
messages = ['389114', '389156', '389168', '389337', '389340', '389344', '392116']
nosy_count = 5.0
nosy_names = ['gvanrossum', 'rhettinger', 'Mark.Shannon', 'v2m', 'BTaskaya']
pr_nums = ['24936']
priority = 'normal'
resolution = None
stage = 'patch review'
status = 'open'
superseder = None
type = None
url = 'https://bugs.python.org/issue43563'
versions = ['Python 3.10']

Linked PRs

• gh-87729: add LOAD_SUPER_ATTR instruction for faster super() #103497
• gh-87729: specialize LOAD_SUPER_ATTR_METHOD #103809
• gh-87729: improve hit rate of LOAD_SUPER_ATTR specialization #104270
• gh-87729: add LOAD_SUPER_ATTR to 3.12 What's New #105125
• [3.12] gh-87729: add LOAD_SUPER_ATTR to 3.12 What's New (GH-105125) #105143

[vladima]

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

[markshannon]

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?

[rhettinger]

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().

[vladima]

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.

[gvanrossum]

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.

[markshannon]

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?

[vladima]

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