Inline bytecode caches

[brandtbucher]

BPO	46841
Nosy	@markshannon, @corona10, @pablogsal, @brandtbucher, @neonene, @iritkatriel, @penguin-wwy
PRs	bpo-46841: Use *inline* caching for BINARY_OP #31543 bpo-46841: Add (undocumented) _co_quickened attribute for code object. #31552 bpo-46841: Move the cache for LOAD_GLOBAL inline. #31575 bpo-46841: Use inline caching for UNPACK_SEQUENCE #31591 bpo-46841: Inline cache for BINARY_SUBSCR. #31618 bpo-46841: Use inline caching for COMPARE_OP #31622 bpo-46841: Use inline caching for attribute accesses #31640 bpo-46841: Improve the failure stats for COMPARE_OP #31663 bpo-46841: Fix error message hacks in GET_AWAITABLE #31664 bpo-46841: Fix BINARY_OP's handling of inline caches #31671 bpo-46841: Use inline caching for calls #31709 bpo-46841: Don't use an oparg counter for STORE_SUBSCR #31742 bpo-46841: Add a _Py_SET_OPCODE macro #31780 bpo-46841: Update adaptive.md for inline caching #31817 bpo-46841: Quicken code in-place #31888 bpo-46841: Don't scan backwards in bytecode. #31901 bpo-46841: Don't jump during throw()  #31968 gh-90997: bpo-46841: Disassembly of quickened code #32099 bpo-46841: remove no-longer-used macro UPDATE_PREV_INSTR_OPARG #32100 bpo-46841: Use a bytes object for _co_code_adaptive #32205 bpo-46841: Avoid unnecessary allocations in code object comparisons #32222

^{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/brandtbucher'
closed_at = None
created_at = <Date 2022-02-24.02:17:14.449>
labels = ['interpreter-core', 'deferred-blocker', '3.11', 'performance']
title = 'Inline bytecode caches'
updated_at = <Date 2022-04-01.11:29:28.677>
user = 'https://github.com/brandtbucher'

bugs.python.org fields:

activity = <Date 2022-04-01.11:29:28.677>
actor = 'Mark.Shannon'
assignee = 'brandtbucher'
closed = False
closed_date = None
closer = None
components = ['Interpreter Core']
creation = <Date 2022-02-24.02:17:14.449>
creator = 'brandtbucher'
dependencies = []
files = []
hgrepos = []
issue_num = 46841
keywords = ['patch']
message_count = 29.0
messages = ['413875', '413901', '413987', '413988', '414188', '414194', '414261', '414277', '414343', '414349', '414358', '414359', '414364', '414371', '414372', '414378', '414391', '414461', '414490', '414523', '414547', '414696', '414763', '414955', '415307', '415676', '415969', '416482', '416483']
nosy_count = 7.0
nosy_names = ['Mark.Shannon', 'corona10', 'pablogsal', 'brandtbucher', 'neonene', 'iritkatriel', 'penguin_wwy']
pr_nums = ['31543', '31552', '31575', '31591', '31618', '31622', '31640', '31663', '31664', '31671', '31709', '31742', '31780', '31817', '31888', '31901', '31968', '32099', '32100', '32205', '32222']
priority = 'deferred blocker'
resolution = None
stage = 'patch review'
status = 'open'
superseder = None
type = 'performance'
url = 'https://bugs.python.org/issue46841'
versions = ['Python 3.11']

Linked PRs

• GH-90997: Shrink the LOAD_GLOBAL caches #102569

[brandtbucher]

...as discussed in faster-cpython/ideas#263.

My plan is for this initial PR to lay the groundwork, then to work on porting over the existing opcode caches one-by-one. Once that's done, we can clean up lots of the "old" machinery.

[markshannon]

We need to decide what to do about dis.

I don't think we should have a show_cache option, as the caches are meaningless junk without quickening (maybe we should drop the CACHE opcode, and just use zeroes).

Instead we should have a show_quickened option, to show the quickened form, which we need to make clear is very much implementation defined.
E.g. Cinder might show the machine code as well.

That way, we can present the cache information as extra data on the quickened form, rather than junk instructions.

[markshannon]

New changeset 0f41aac by Brandt Bucher in branch 'main':
bpo-46841: Use inline caching for BINARY_OP (GH-31543)
0f41aac

[markshannon]

Making this a release blocker, as we really cannot leave this half finished for the release.

Shouldn't be a problem, as we'll have it done in a week or so.

[markshannon]

New changeset 424ecab by Brandt Bucher in branch 'main':
bpo-46841: Use inline caching for UNPACK_SEQUENCE (GH-31591)
424ecab