Adds __hash__ method to @dataclasses.dataclass, refs #11463

[sobolevn]

Closes #11463
Closes #11495

I still need to add a test case for 3.9, where there's no unsafe_hash arg, but __hash__ is still added somehow.
Right now I just want to see the CI results.

[JukkaL]

Thanks for the PR! I like that all the possible cases are carefully tested. Left some minor comments, overall looks good.

[JukkaL]

I think that this makes the conditions harder to understand. If you think that the conditions would be too verbose otherwise, maybe instead rename unsafe_hash to something shorter, such as unsafe. Now you could write something like if not unsafe and eq and not frozen.

[sobolevn]

if ((not unsafe_hash and not eq and not frozen)
            or (not unsafe_hash and not eq and frozen)):

I think that both are quite hard to read 😞
Mine was inspired by pattern matching technique I use in different functional languages.

Anyways, I've changed it! 👍

[JukkaL]

The mypy_primer failure seems to indicate a real issue. For example, this generates a false positive:

from dataclasses import dataclass

a = f()  # Forward ref to force two semantic analysis passes

@dataclass(unsafe_hash=True)
class C:  # Cannot overwrite attribute "__hash__" in class "C"
    x: str

def f(): pass

It looks like the new logic doesn't work correctly if we need multiple semantic analysis passes. If the previous pass had added a __hash__ method, we shouldn't complain about it in the second pass.

[sobolevn]

Let's try this solution. Now I check that given variable was not plugin generated.

[JukkaL]

Can you fix the conflicts? I can do another round of review once the PR is up-to-date.

[sobolevn]

I somehow nuked my old fork and that's why all my old PRs are now closed.
Thanks a lot to @AlexWaygood for letting me know.

I can recreate some of them. Here's a list of things I still want to get merged:

• Adds __hash__ method to @dataclasses.dataclass, refs #11463 #11496
• Disallow to use TypeGuard type outside of .ret_type, refs #11765 #11790
• @classmethod and @staticmethod were not shown in error messages when cached #11793
• Fixes how ** affects inference in {} expressions, refs #11691 #11693
• Fix @dataclass(frozen=True) fail on name shadowing #12088

[AlexWaygood]

I can recreate some of them. Here's a list of things I still want to get merged:

I was a big fan of these, as well :)

• Do not count __slots__ as a protocol member #11885
• Adds Literal type math #11992

[github-actions]

According to mypy_primer, this change has no effect on the checked open source code. 🤖🎉