Clumsy dispatching on interpreter entry.

[markshannon]

BPO	44878
Nosy	@markshannon, @neonene
PRs	bpo-44878: Add missing DISPATCH() #27715 bpo-44878: Further ceval entry cleanup #27725 bpo-44878: Remove the switch from the main interpreter loop when using computed gotos. #27726 bpo-44878: Remove loop from interpreter. All dispatching is done by gotos. #27727

^{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 2021-08-12.10:54:43.596>
created_at = <Date 2021-08-10.13:42:38.715>
labels = ['3.11', 'performance']
title = 'Clumsy dispatching on interpreter entry.'
updated_at = <Date 2021-08-18.09:37:17.120>
user = 'https://github.com/markshannon'

bugs.python.org fields:

activity = <Date 2021-08-18.09:37:17.120>
actor = 'Mark.Shannon'
assignee = 'Mark.Shannon'
closed = True
closed_date = <Date 2021-08-12.10:54:43.596>
closer = 'Mark.Shannon'
components = []
creation = <Date 2021-08-10.13:42:38.715>
creator = 'Mark.Shannon'
dependencies = []
files = []
hgrepos = []
issue_num = 44878
keywords = ['patch']
message_count = 6.0
messages = ['399324', '399383', '399391', '399440', '399823', '399826']
nosy_count = 2.0
nosy_names = ['Mark.Shannon', 'neonene']
pr_nums = ['27715', '27725', '27726', '27727']
priority = 'normal'
resolution = 'fixed'
stage = 'resolved'
status = 'closed'
superseder = None
type = 'performance'
url = 'https://bugs.python.org/issue44878'
versions = ['Python 3.11']

[markshannon]

On entering the interpreter (_PyEval_EvalFrameDefault) we need to check for tracing in order to record the call.
However, we don't do this cleanly resulting in slow dispatch to the non-quickened instruction on every call/next.

[markshannon]

New changeset 3f3d5dc by Mark Shannon in branch 'main':
bpo-44878: _PyEval_EvalFrameDefault readability improvements (GH-27725)
3f3d5dc

[markshannon]

New changeset f66d00f by Mark Shannon in branch 'main':
bpo-44878: Remove the switch from the main interpreter loop when using computed gotos. (GH-27726)
f66d00f

[markshannon]

New changeset a530a95 by Mark Shannon in branch 'main':
bpo-44878: Remove loop from interpreter. All dispatching is done by gotos. (GH-27727)
a530a95

[neonene]

FYI, PR27727 ("Remove loop...") seems to be a bit slower than the previous commit (f08e6d1) on my Windows build (msvc14.29.16.10). pyperformance shows that

Windows x64 PGO: 34 slower, 11 faster, 13 not significant, Geometric mean: 1.02x slower
Windows x86 PGO: 28 slower, 17 faster, 13 not significant, Geometric mean: 1.02x slower

Undoing PR27727 on current cpython-main branch also get speed-ups by 1-2% on average.

[markshannon]

I'm somewhat surprised by that. After all, the only change in control flow was the change from a break to a goto in exception handling.
I would have expected PR27726 to have made much more difference.

There are a few possibilities, including:

1. It's just a random fluctuation from tiny changes in alignment.
2. MSVC aligns loops, but not switches, or vice-versa, and that makes a systematic difference.

I suspect that it is (1), but this is a bit worrying nonetheless.