0.920 Regression: literal unions aren't inferred from conditionals anymore

[elprans]

test.py:

def foo(h: bytes) -> int:
    if h[5] == 2:
        byteorder = "big"
    elif h[5] == 1:
        byteorder = "little"
    else:
        raise AssertionError(
            f"unexpected ELF endianness: {h[5]}"
        )
    return int.from_bytes(h[16:18], byteorder=byteorder, signed=False)

Results in

$ mypy test.py
1.py:10:17: note: Revealed type is "builtins.str"
1.py:11:47: error: Argument "byteorder" to "from_bytes" of "int" has incompatible type "str"; expected "Union[Literal['little'], Literal['big']]"

With mypy-0.920 and later. 0.910 correctly inferred byteorder as Literal['little'] | Literal['big']

[pranavrajpal]

I don't think mypy ever inferred the type of byteorder as Literal['little', 'big']. I think we always inferred str but 0.920 changed the type signature of int.from_bytes to take a Literal instead of a str, turning this into an error. (That change was added in python/typeshed#5715 and added to mypy in #10862, which was included in 0.920).

You can silence the error by adding an explicit Literal["little", "big"] type for byteorder.

[hauntsaninja]

TBH I find several of these typeshed Literal changes to be quite problematic, e.g. here's one I recently reverted python/typeshed#6566

[elprans]

@pranavrajpal, you're right, 0.910 also doesn't infer byteorder as Literal. I somehow convinced myself that it did. Not an inference regression then. Thanks for the explanation!