bpo-43224: Forbid TypeVar substitution with Unpack

[serhiy-storchaka]

https://bugs.python.org/issue43224

[gvanrossum]

Remind me where PEP 646 forbids this? Assuming Unpack[Ts] === *Ts, it seems this is forbidding the following:

class G(Generic[*Ts]): ...
B = G[T, *Ts, str, T2]
X = B[int, *Ts]  # <-- TypeError here

What is wrong with X that isn't wrong with B?

[serhiy-storchaka]

I'll rewrite it as B[int, *Ts2], so we could unambiguously refer to variables.

AFAIK PEP 646 does not cover such case. It does not specify the result.

But what can be the result of substitution B[int, *Ts2]? T is substituted with int, T2 should be substituted with the last item of *Ts2, and *T should be substituted with all but the last items of *Ts2. But we cannot express this, because Ts2 is a variable. If we substitute *Ts2 with an empty sequence of types (X[()]), there would not be a value for T2.

I think that it is better to make it an error:

def f(a: tuple[*Ts2]) -> B[int, *Ts2]: ...

You should write

def f(a: tuple[*Ts2, T3]) -> B[int, *Ts2, T3]: ...

to make it having some sense.

There is a workaround of this problem. It may even make the code more clear. You can see, that a should be a tuple containing at least 1 item. T2 in B will be substituted with the type of the last item, and *Ts in B will be substituted with the rest of types.

[JelleZijlstra]

@mrahtz what do you think of this case?

[mrahtz]

Thanks for the prod, Jelle. I agree with Serhiy - looks like this is one of the cases we should forbid at runtime.

[JelleZijlstra]

Looks good in principle, pending resolution of Guido's question.