Home Prefer views::meow

Prefer views::meow

With the adoption of Ranges in C++20 (and with a lot of new additions in C++23), we have a lot of new, composable, range-based algorithms at our disposal. There are, however, several ways to spell the usage of such algorithms:

std::vector<int> v = {1, 2, 3};
auto square = [](int i){ return i * i; };

auto a = v | std::views::transform(square);
auto b = std::views::transform(v, square);
auto c = std::ranges::transform_view(v, square);

Which should you prefer?

The answer is either a or b, never c.

The use of std::views::transform here is valid, you do not have to write out std::ranges::views::transform. The standard library provides std::views as a namespace alias for std::ranges::views - as you can see in the synopsis for <ranges> (towards the end).

views::meow is the algorithm

views::meow is the actual, user-facing algorithm. It takes a viewable_range, and possibly some other arguments, and produces a new view whose behavior meets the goal of the algorithm. In the above example, views::transform takes a range (v) and a unary function (square) and produces a view whose elements are the result of applying that function to every element of the original range.

The use of meow here is as a metasyntactic variable. Other people might use foo or some other more obvious placeholder like XXX. I somehow picked up this practice from STL and Tim Song, even though my avatar for this blog is a dog.

In the typical case, the way to implement views::meow is by creating a new type, ranges::meow_view, that wraps the original view and those other arguments and has the expected behavior. For an algorithm like transform, this is quite simply what it does:

Given subexpressions E and F, the expression views​::​transform(E, F) is expression-equivalent to transform_­view(E, F).

But this isn’t actually the case for all range adaptors, and it need not even be the case for views::transform.

Not all wrapping is necessary

Consider one of the new C++23 adaptors: views::as_const. The job of views::as_const(E) (or E | views::as_const, if you prefer), is to produce a range whose elements you cannot mutate. If E were a range of int&, then views::as_const(E) would give you a range of int const&. It does so by returning a ranges::as_const_view(E).

But what if E were already a range of int const&? In that case, E is already a constant view - we don’t need to do any work to produce a constant view out. So views::as_const(E) can simply return… E. ranges::as_const_view(E) cannot do that, since that’s a type - but views::as_const(E) can, since it’s an algorithm. It can make smarter, more efficient choices.

The same holds for views::as_rvalue(E) (which simply propagates E if it’s already a range of rvalues) and views::common(E) (which likewise propagates E if it is already a common view). In these cases, the wrapping may not even be valid - views::common(E) gives you back a common view regardless of whether E is common or not, but ranges::common_view(E) requires E to be non-common.

More efficient implementations

In other cases, it is possible for an algorithm to produce a more efficient implementation - where efficient isn’t just a runtime behavior, it’s also a compile-time one. Instantiating fewer templates, with less wrapping, means faster compile time - and likely better optimization outcomes, since there’s less work to do.

There are several cases of this in the standard library as well:

void f(span<int> s) {
    auto a = s | views::take(1);
    auto b = s | views::drop(1);
    auto c = s | views::as_const;

In this example, ranges::take_view(s, 1), ranges::drop_view(s, 1), and ranges::as_const_view(s) would all be valid ranges, that all have the desired behavior of the algorithm. They are all semantically correct. But they all require further instantiations, with more overhead (that is hopefully optimized out).

But that’s not what views::take, views::drop, and views::as_const do. Instead, a and b are also actually objects of type span<int> (constructed appropriately) and c is a span<int const>. That’s strictly better than the alternative.

Likewise, consider:

auto r = /* some bidirection range */;

auto rev1 = views::reverse(r);
auto rev2 = views::reverse(rev1);

rev2 is r, reversed twice. views::reverse(E) recognizes when E is itself a reversed view, and short-circuits - rev2 is simply r (or, more precisely, views::all(r)). But had we used ranges::reverse_view(r) and ranges::reverse_view(rev1), that’s not the behavior we’d get - instead we’d have ended up with a ranges::reverse_view<ranges::reverse_view<V>>. Double-wrapped, instead of not-wrapped.

Even in the original transform example, it is possible to do better:

r | views::transform(f) | views::transform(g)

You could imagine that an implementation could recognize that it’s adapting a transform-ed range, and internally compose the two functions - so that internally this becomes:

r | views::transform(f >> g)

That’s not what views::transform does in C++20, and indeed it is explicitly specific to not do that. But if you just use views::transform, then perhaps it could in the future.

ranges::meow_view is an implementation detail

The right way to think about ranges::meow_view is that it’s simply an implementation detail of views::meow. Sometimes the latter gives you the former, sometimes it gives you something else. But you should really think about views::meow as simply giving you something that satisfies the semantics of the algorithm. views::meow also supports piping, so it’s just more convenient to begin with.

And lastly, views::meow(E) is just shorter than ranges::meow_view(E). So even if there weren’t several compelling reasons to prefer it already, it’s also less to type. So if I haven’t convinced you yet, there’s also this.

The question might be, at this point, why do we even specify ranges::meow_view to begin with? And there, I think, the answer simply has to do with complexity. I think users should consider these types as exposition-only, but it is helpful for the implementations to have all their behavior specified. A lot of it is subtle. Plus a lot of adaptors expose their internals (via base() or pred(), etc.) so it’s not just a question of views::transform(E, F) yielding a range with particular semantics, it’s also a question of how else you can interact with the result. As tedious as it is to provide these specifications, I do have to grudgingly accept that they provide some value.

Surely there’s at least some use for explicit ranges::meow_view?

There is precisely two situations where it makes sense to explicitly use ranges::meow_view over views::meow. The obvious one is: when you’re implementing views::meow. But that’s a boring answer.

The more interesting answer is: when you’re implementing a different view. One example here might be zip_transform. views::zip_transform(F, Es...) is a range that whose elements are the result of F(es...) for all the elements es... of Es.... This is quite closely related to zip, and indeed is very similar to:

views::zip(Es...) | views::transform(hof::unpack(F))

(see here for hof::unpack).

How would you implement views::zip_transform? Fundamentally, you’re still iterating over all the ranges at the same time - that’s still a zip. And, indeed, the easiest way to implement this is to, internally, use ranges::zip_view (as you can see here):

template <move_constructible F, input_range... Views>
    requires /* bunch of other stuff */
class zip_transform_view : public view_interface<zip_transform_view<F, Views...>> {
    movable_box<F> fun_;
    zip_view<Views...> zip_;

Although since this is the standard library, we don’t just dereference zip_.begin() (producing a std::tuple) and then use std::apply on that result - instead we sidestep the construction of the std::tuple and get access to the underlying std::tuple<iterator_t<Views>...> directly.

But the point here is - having a member zip_view makes it easier to implement zip_transform_view.

If you’re not implementing a view, there’s probably no other reason to explicitly write ranges::meow_view.


Always prefer views::meow over ranges::meow_view, unless you have a very explicit reason that you specifically need to use the latter - which almost certainly means that you’re in the context of implementing a view, rather than using one.

This post is licensed under CC BY 4.0 by the author.