Also, the current CI failures will have to be addressed of course. My comments above were just concerns that I noticed at a glance. I suspect that some of the failures may have been occurring as a result of theseld._buffer code typos that I mentioned, but I haven't looked over them yet.

@aeros Thanks for your comments, I've hopefully fixed the code / logic mistakes.

@tzickel

Thanks, but it's a small one-time overhead instead of a small overhead on each object initialization

I wasn't referring to repeating thegetattr(socket.socket, "sendmsg", False) every time the object is created; that's not necessary. Here's an example of what I had in mind:

Current:

sendmsg=getattr(socket.socket,"sendmsg",False)# [snip]defwritelines(self,lines):ifnotsendmsg:returnself.write(b''.join(lines))# [snip]

Recommended:

# Used to determine if platform supports sendmsg to sockets# [_NAME is our typical convention for internal globals, to differentiate it]_SENDMSG=Noneclass_SelectorSocketTransport(_SelectorTransport):# [snip]def__init__(self,loop,sock,protocol,waiter=None,extra=None,server=None):global_SENDMSG# [ensures the `getattr()` is only done once, but doesn't create extra# overhead when select_events is imported]if_SENDMSGisNone:_SENDMSG=getattr(socket.socket,"sendmsg",False)self._sendmsg=_SENDMSG# [snip]defwritelines(self,lines):ifnotself._sendmsg:returnself.write(b''.join(lines))# [snip]

Does that make it more clear?

Also, as a side note, there's typically no need to force push over old commits for PR branches in the CPython repo since we squash prior to merging. It canoccasionally help if it the history becomes muddled over time, but it can also make changes made to the PR harder to follow for review purposes.

Seehttps://discuss.python.org/t/pep-601-forbid-return-break-continue-breaking-out-of-finally/2239/42. It's also mentioned at the end ofhttps://devguide.python.org/pullrequest/#submitting.

#31871 is the newer resurrection of the idea.