Remove the usage of the C stack in Python to Python calls

[pablogsal]

BPO	45256
Nosy	@gvanrossum, @pmp-p, @markshannon, @serhiy-storchaka, @zooba, @ctismer, @corona10, @pablogsal, @brandtbucher, @Fidget-Spinner, @thesamesam
PRs	bpo-45256: Remove the usage of the C stack in Python to Python calls #28488 bpo-45256: Small cleanups for the code that inlines Python-to-Python calls in ceval.c #28836 bpo-45256: Fix cleanup of stolen locals for Python-to-Python calls #28905 bpo-45256: Avoid C calls for more Python to Python calls. #28937 bpo-45256: Rationalize code around Python-to-Python calls a bit. #29235 bpo-45256: Don't call at the C level when calling bound-methods from Python code. #29238 bpo-45256: Don't track the exact depth of each InterpreterFrame #30372

^{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 = <Date 2022-01-06.12:52:24.256>
created_at = <Date 2021-09-21.10:41:59.834>
labels = ['3.11']
title = 'Remove the usage of the C stack in Python to Python calls'
updated_at = <Date 2022-01-06.12:52:24.256>
user = 'https://github.com/pablogsal'

bugs.python.org fields:

activity = <Date 2022-01-06.12:52:24.256>
actor = 'Mark.Shannon'
assignee = 'none'
closed = True
closed_date = <Date 2022-01-06.12:52:24.256>
closer = 'Mark.Shannon'
components = []
creation = <Date 2021-09-21.10:41:59.834>
creator = 'pablogsal'
dependencies = []
files = []
hgrepos = []
issue_num = 45256
keywords = ['patch']
message_count = 21.0
messages = ['402311', '402429', '402430', '402485', '402491', '402504', '402518', '402546', '402763', '402766', '402767', '402780', '402785', '403538', '403543', '403832', '404166', '405109', '405193', '406462', '409753']
nosy_count = 11.0
nosy_names = ['gvanrossum', 'pmpp', 'Mark.Shannon', 'serhiy.storchaka', 'steve.dower', 'Christian.Tismer', 'corona10', 'pablogsal', 'brandtbucher', 'kj', 'thesamesam']
pr_nums = ['28488', '28836', '28905', '28937', '29235', '29238', '30372']
priority = 'normal'
resolution = 'fixed'
stage = 'resolved'
status = 'closed'
superseder = None
type = None
url = 'https://bugs.python.org/issue45256'
versions = ['Python 3.11']

Linked PRs

• gh-89419: gdb: fix printing AttributeError in py-locals when no frame is available #100611
• [3.11] gh-89419: gdb: fix bug causing AttributeError in py-locals when no frame is available (GH-100611) #100738

[pablogsal]

Removing the usage of the C stack in Python-to-Python calls will allow future optimizations in the eval loop to take place and can yield some speed ups given that we will be removing C function calls and preambles by inlining the callee in the same eval loop as the caller.

[serhiy-storchaka]

AFAIK Mark Shannon proposed this idea, but it was rejected.

[pablogsal]

What was rejected was https://www.python.org/dev/peps/pep-0651/ which included this idea but had a lot more stuff in it. In particular, it was rejected because it gave semantics to overflow exceptions (two exceptions were proposed), new APIs and it had lack consistent guarantees for different platforms, among other considerations.

This version is just for optimization purposes with no changes in semantics.

[ctismer]

Hey guys, you know that you are about to implement the core idea of Stackless Python, right? :-D

[markshannon]

I've trying to do this since about 2011 :)

[zooba]

I fully support implementing the core idea of Stackless Python :)

I spent a whole EuroPython a couple of years back discussing the idea with (apparently) everyone except Mark.

Though I wouldn't like to lose the ability to extract the Python stack by inspecting native memory alone. That is very useful for debugging, particularly when you've only got a crash dump. So provided all the code objects are only an indirection or two away from a local (or better yet, parameter) value, it should be fine.

[pablogsal]

Though I wouldn't like to lose the ability to extract the Python stack by inspecting native memory alone.

Don't worry about it, I am personally making sure that keeps being possible. it will need some changes in the tools, but not any more that any other changes between minor versions of the interpreter

[ctismer]

FYI., in Stackless Python I built a chain of frames by a double linked list. This was the replacement for the current frame/stack mix.
Debugging was still quite easy, following this frame chain.

Note that it is a rather easy step to add the capability to squirrel the whole current chain away and replace it by another one. Add some header to such a chain, and you can call it "Tasklet".

[markshannon]

PR 28488 has no NEWS entry, or What's New entry.

However, adding multiple entries will be confusing, so that's best left until all calls to Python functions and method don't use the C stack.

[ctismer]

FWIW, getting all function to avoid the C stack will most probably take a long time, if it happens at all. Especially functions which might call into Python multiple times must be re-designed heavily, turned into multiple pieces to be in tail position. Been there, long ago... :)

[pablogsal]

We are starting to optimize first the easy cases of CALL_FUNCTION and move slowly to add more and more calls into it. This approach has the advantage that allow us to take it in small increments.

[ctismer]

Very much appreciated approach.
Too bad that things stop when people are writing extensions as usual. Or do you think we can teach them how to avoid the C stack? Anyway, good luck!

[zooba]

The goal is reduced stack depth, not reframing the entire call model around not having a C stack.

We can't even reasonably rewrite getattr() without supporting callbacks from C into Python, so further generalisation is very unlikely.

But if you inspect the native stack of most Python programs, you'll see that it's mostly taken up with calls within Python code. Compressing all of those is a significant advantage, akin to inlining the Python code at compile time, even if it doesn't see "through" native calls.

[pablogsal]

New changeset b4903af by Pablo Galindo Salgado in branch 'main':
bpo-45256: Remove the usage of the C stack in Python to Python calls (GH-28488)
b4903af

[pablogsal]

New changeset 543acbc by Pablo Galindo Salgado in branch 'main':
bpo-45256: Small cleanups for the code that inlines Python-to-Python calls in ceval.c (GH-28836)
543acbc

[markshannon]

New changeset 3901c08 by Pablo Galindo Salgado in branch 'main':
bpo-45256: Fix cleanup of stolen locals for Python-to-Python calls (GH-28905)
3901c08

[markshannon]

New changeset 70945d5 by Mark Shannon in branch 'main':
bpo-45256: Avoid C calls for most Python to Python calls. (GH-28937)
70945d5

[pablogsal]

Unfortunately, seems that #28937 has broken the AMD64 FreeBSD Shared 3.x buildbot:

https://buildbot.python.org/all/#/builders/483/builds/1003/steps/5/logs/stdio

The buildbot was green until we merged this

[markshannon]

New changeset 7f61d9d by Mark Shannon in branch 'main':
bpo-45256: Rationalize code around Python-to-Python calls a bit. (GH-29235)
7f61d9d

[markshannon]

https://bugs.python.org/issue45829 is the related issue for special methods

[markshannon]

New changeset 332e6b9 by Brandt Bucher in branch 'main':
bpo-45256: Don't track the exact depth of each InterpreterFrame (GH-30372)
332e6b9