Nevin ":-)" Liber

Just to be clear, this new example is not about compatible types (as there is only one type of `Map`).

Hashed maps like `UnorderedMap` cannot make ordering guarantees about where elements go because of hash collisions. This is why, for instance, `std::unordered_map::operator==` can have an O(N^2) run time in the...

Suppose the hash of distinct `int` keys `z` and `y` collide. Given the following code: ``` using UM = UnorderedMap; UM a; auto inserteday = a.insert(y); auto insertedaz = a.insert(z);...

Then what do you mean by "internal state"?

I think what we are doing now (just copying the internal state) is the right way to go. I don't think we should be turning deep_copy into something O(N) or...

I don't know that either `MDSPAN_DRAFT_NAMESPACE` or `MDSPAN_PROPOSED_NAMESPACE` should be defaulted to `std::experimental`. There aren't going to be any more Library Fundamental TSes (after the current v3 is voted in)...

I'm not sure if having `MDSPAN_EXTENSION_NAMESPACE` default to `mdspan` is a good idea (from the user perspective), because that is also the name of the class. At a minimum, it'll...

`Sum` already uses `reduction_policy::sum()` when calling `init(value_type&)`. For `Sum`, we can fall back on a default constructed `value_type`, such as in #7167. Open questions: - Should we? I think it...

> @nliber This needs to be addressed in the `parallel_scan` implementation, not in the `Sum` implementation (since that isn't used in the backend implementations). The latter is a separate discussion....

> > Should `init(value_type&)` implementations be `static`? I didn't see any dependent on member variables. > > If we call it as a static member function internally that might break...