Thanks for all the mypyc PRs! I will probably leave them all to @JukkaL and @p-sawicki (or maybe I will take one-two if they are busy).

Thanks for the! This is a useful improvement. I found a few regressions when trying this out. First, this crashes the compiler:

def f(x: object, classes: tuple[type, ...]) -> bool:
    return isinstance(x, (int, *classes))

The assertion here unexpectedly fails (this is pretty esoteric though):

def f(x: object) -> bool:
    cls: type = int
    return isinstance(x, (cls, (cls := str)))

assert f(1)

Maybe fall back to the old approach in these cases? However, this could slightly change semantics.

@JukkaL Oof. Good catch. I'll be honest, out of all the mypyc PRs I opened this weekend, this one was the one I had the least confidence in. Fwiw, #22099 is second...

You are correct, I think the only way to handle those is to fall back to the old method since a star-expansion can have a variable number of items of unreliable types, and a walrus operator can change the values. I think the walrus regression would be the mypyc equivalent of #21249 . I've also added tests for this.

Also just testing now I think I found an issue that also fails on master which is similar to the second regression. I don't see it reported anywhere. Your test there is about when the walrus replaces a var in the tuple, but I seem to have problems when it replaces another arg as well. For example, pair(i, (i := 1)) passes (1, 1) instead of (0, 1) on master.