Create frame objects lazily when needed

[markshannon]

BPO	44590
Nosy	@gvanrossum, @ncoghlan, @vstinner, @markshannon, @pablogsal
PRs	bpo-44590: Lazily allocate frame objects #27077 WIP bpo-44800: Rename _PyInterpreterFrame to _Py_framedata #27525

^{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 = 'https://github.com/markshannon'
closed_at = <Date 2021-07-29.10:31:08.811>
created_at = <Date 2021-07-09.10:53:40.343>
labels = ['interpreter-core', '3.11', 'performance']
title = 'Create frame objects lazily when needed'
updated_at = <Date 2022-02-23.15:14:23.151>
user = 'https://github.com/markshannon'

bugs.python.org fields:

activity = <Date 2022-02-23.15:14:23.151>
actor = 'vstinner'
assignee = 'Mark.Shannon'
closed = True
closed_date = <Date 2021-07-29.10:31:08.811>
closer = 'Mark.Shannon'
components = ['Interpreter Core']
creation = <Date 2021-07-09.10:53:40.343>
creator = 'Mark.Shannon'
dependencies = []
files = []
hgrepos = []
issue_num = 44590
keywords = ['patch']
message_count = 4.0
messages = ['397196', '398220', '398696', '413798']
nosy_count = 5.0
nosy_names = ['gvanrossum', 'ncoghlan', 'vstinner', 'Mark.Shannon', 'pablogsal']
pr_nums = ['27077', '27525']
priority = 'normal'
resolution = 'fixed'
stage = 'resolved'
status = 'closed'
superseder = None
type = 'performance'
url = 'https://bugs.python.org/issue44590'
versions = ['Python 3.11']

[markshannon]

In https://bugs.python.org/issue44032 we moved most of the data in the frame stack, that is the locals, stack, and "specials" (globals, builtins, code etc), from the heap allocated stack to a (mostly) contiguous array in memory.
That offered some speed up due to better cache locality, and faster allocation of frame objects (they are now fixed size so can use a simple freelist).

However, data is still split between the stack allocated frame and the heap allocated frame.

We should move the remaining data to the stack and only allocate heap objects lazily when needed for tracebacks and the like. This should improve performance further by removing the vast majority of heap frame allocations and further improving locality of reference.

Not only does this have immediate performance benefits, it also paves the way for even better Python-to-Python calls by allowing stack frames to overlap and pass arguments with the minimal amount of copying and INCREF/DECREF pairs.

[markshannon]

New changeset ae0a2b7 by Mark Shannon in branch 'main':
bpo-44590: Lazily allocate frame objects (GH-27077)
ae0a2b7

[ncoghlan]

The newly linked pull request isn't actually for this ticket, it's for bpo-44800, a follow-up refactoring proposal related to the variable, struct field, and API naming schemes used for the new lighter weight execution frames.

However, the commit message and PR description refer back to this ticket as well, so the RoundUp automation picked it up.

[vstinner]

I created bpo-46836: "[C API] Move PyFrameObject to the internal C API".