[WIP] handle diverging protocols

[ilevkivskyi]

No description provided.

[ilevkivskyi]

@JukkaL I am still investigating things here, but there is something interesting I noticed: pandas requires insanely large protocol stack (it still crashes with stack of size 20). I am not sure how exactly, but this is caused by the cycle we have is_subtype() -> unify_generic_callable() -> infer_constraints() -> make_simplified_union() -> is_subtype(). That make_simplified_union() is technically needed to make type inference stable w.r.t. equivalent types.

But, the fun part is removing union simplification there makes pandas (and various other libraries that depend on it) ~2x faster (yes, you read that right, twice faster, not 2% faster). This however makes couple tests fail and causes a bit of primer fallout (mostly because of redundant unions in some asyncio stubs).

For now I will probably simply set some high protocol stack size. This should fix at least three crashes. But I also think we should consider cutting some corners in type inference (in a separate PR), since 2x speed-up looks very tempting.

[github-actions]

According to mypy_primer, this change doesn't affect type check results on a corpus of open source code. ✅

[ilevkivskyi]

OK, I think I have enough data to make a real PR, I will close this one an open a new clean PR.