bpo-33332: Add signal.valid_signals()

[pitrou]

https://bugs.python.org/issue33332

[vstinner]

It would be interesting to also implement the feature on Windows where "valid signals" is hard to guess for a Unix user. See signal_signal() of Modules/signalmodule.c for valid signals: I count 6 signals, 7 if SIGBREAK is also available.

[vstinner]

What do you think of testing len(s) < NSIG? Maybe len(s) <= NSIG?

[pitrou]

I'm not sure it's very useful.

[vstinner]

The purpose is to test if NSIG is consistent with sigfillset(), to check if something is really wrong on a platform. I would like to know if range(0, NSIG) miss some signals on some platforms :-)

[vstinner]

You may also modify:

• Lib/asyncio/unix_events.py: "if not (1 <= sig < signal.NSIG):" in _check_signal() -- don't forget to patch Lib/test/test_asyncio/test_unix_events.py
• Lib/test/support/__init__.py: SaveSignals uses "self.signals = list(range(1, signal.NSIG))"

[pitrou]

Thanks for the comments. You're right about Windows. I'll also fix asyncio and test.support.

[1st1]

Big +1 for this feature. Can you make it return a frozenset though?

[pitrou]

Frozen sets are nice if you want to hash them, otherwise I don't really see the point.

[1st1]

Frozen sets are nice if you want to hash them, otherwise I don't really see the point.

What's the point of being able to modify the returned set? ;) If you return a frozenset you can cache it (something that I would recommend doing in this PR anyways).

This isn't a deal breaker, it's just that I would use a frozen set here. Overall LGTM.

[pitrou]

Other functions like signal.pthread_sigmask and signal.sigpending return plain sets. So we could return a frozenset from signal.valid_signals but that sounds a bit gratuitous. @vstinner, what do you think?

[vstinner]

First I wanted to get a list, but then I read the PR and I like reusing existing code :-)

If someone wants an immutable type, it's very simple: frozenset(signal.valid_signals()).

An immutable type might be justified if it was a module constant, but it's a function which creates a new object at each call, so IMHO a set is just fine.

I don't want to modify other signal functions just to be pedantic. I like Antoine's rationale to use set in valid_signals().