New semantic analyzer: Support idem forward aliases

[ilevkivskyi]

This is a follow-up for #6390

This tricky corner case is used in some real code and is supported by old semantic analyzer

C = C
class C:  # type: ignore
    pass

x: C

But with new analyzer this results in Invalid type since the first definition always wins. Probably we just need to special-case this somehow.

[ilevkivskyi]

EDIT: The new behavior is actually correct, see below.

Another pattern that we need to support with new semantic analyzer is this:

class In:
    ...
class Out:
    In = In  # recognized as variable

this currently (both before and after assignment refactor) causes a bunch of troubles in internal code bases. Raising priority to high.

[ilevkivskyi]

Oh, wait, this is actually correct, by the rules non-indexed assignments at class scope are always variables, so this was previously a false negative, OK back to normal priority.

[ilevkivskyi]

Another repro for this issue with a weird error:

x: C
class C:
    pass
C = C

fails with

main:2: error: Name 'C' already defined on line 4

cc @JukkaL

[JukkaL]

The examples look meaningless, since the code won't work at runtime. I wonder if the actual code where this comes up is different (involving try/except perhaps). This could also affect the way we'd fix this.

[ilevkivskyi]

Yes, the actual code is like

try:
    C = C
except NameError:
    class C:
        ...

Also the second example works without try/except.

I think we should fix this before the release since this is actually used in some projects.

[JukkaL]

Yes, it would be good to make this better. I'm increasing priority.

But even that example seems incomplete, as that will always generate a NameError. I assume it's more like this:

try:
    ModuleNotFoundError = ModuleNotFoundError
except NameError:
    class ModuleNotFoundError(Exception):
        ...

The assignment will fail on some Python versions and pass on some.

[ilevkivskyi]

Most likely, I don't remember TBH, also aiohttp (who first reported this) used to have tons of * imports. This may be some other name they define in another module depending on sys.platform and then star import it.

[ilevkivskyi]

I did some digging and it looks like this happens sometimes after star imports from C acceleration modules.