Hot takes on flat_map, resize_default_init, and ranges::to

The pre-Kona WG21 mailing has been out for a little while. Tonight I looked at a few papers; here’s my hot takes.

P0429R6 “A Standard flat_map: I don’t honestly think flat_map is well-baked enough to put into the Standard Library; but at least it’s been in Boost for a while (if I understand correctly) (EDIT: I do not), and it fits into the STL’s scheme of “container adaptors” pretty naturally. The order of the template parameters is awkward but probably correct.

Some minor typos, such as Allocator for Alloc in some of the text.

P0429 does innovate in one area: it has a bunch of constructors taking initializer_list<value_type>&&. This doesn’t do what it sounds like. Consider:

FlatSet(std::initializer_list<T>&& il) {
    for (auto&& elt : il) {

You might think that this “moves-out-of” the elements of the initializer list. That would be efficient, but also horribly incorrect if it actually happened!

for (int i=0; i < 10; ++i) {
    FlatSet<std::string> fs { "abc", "def", "ghi" };
    // ...

If “moving-out-of” initializer_list elements were permitted, then on the second time through this loop, all the string elements of the initializer_list would be in an already-moved-out-of state, and bad stuff would happen. (Yes, initializer_lists are more or less static. The actual semantics are a huge mess. See Jason Turner’s excellent C++Now 2018 talk initializer_lists Are Broken — Let’s Fix Them” for the horrendously gory details.)

Fortunately, initializer_list prevents the naïve programmer from moving-out-of its elements, by forcing its nested reference type to be const T&, not T&. So in our loop in the constructor above,


elt has type const std::string&, which means std::move(elt) has type const std::string&&, which means that emplace_back will end up perfectly forwarding to string’s copy constructor, not its move constructor. Result: taking an initializer_list by non-const rvalue reference is functionally equivalent to taking initializer_list by const lvalue reference, or indeed taking it by value. (“By value” is the single appropriate way to take initializer_list.)

The only thing you accomplish by taking initializer_list by rvalue reference is, you break everyone who thinks they can “perfect-forward” initializer_list by value as a special case. For example, std::make_optional.

auto opt = std::make_optional<FM>(
    // this finds the initializer_list<U> overload of make_optional
        FM::value_type{"1", "abc"},
        FM::value_type{"2", "abc"},

With FM(initializer_list<value_type>), no problem. But with FM(initializer_list<value_type>&&) you get a cryptic error from deep in the bowels of make_optional.

Always pass initializer_list by value.

P1072R3 “basic_string::resize_default_init: Ship it! I’ve watched this one for a while, and I think this is as good as it’s going to get. The library folks have already approved its sister proposal P1020 for make_shared_default_init<int[]>(100) and make_unique_default_init<int[]>(100). Giving the same powers to std::string just makes sense.

P1072 specifically does not propose to add the resize_default_init method to vector (or deque or list or forward_list, all of which have resize methods — did you know?), because it’s unclear whether a general facility would be useful, whereas anyone who works with parsing of character data (e.g. receiving packets from a network socket) knows the use-case for string::resize_default_init.

It would be even awesomer to be able to malloc a buffer, put some data in it, and then hand that buffer over to be managed by a std::string (or even a basic_string with a custom allocator). However, that rabbit hole is very deep and twisty, and I’m glad the authors of P1072 backed away from it.

But read on!

P1206R1 “ranges::to: A function to convert any range to a container”: to is such a great userspace identifier! Please don’t drop it into the black hole of ADL!

The one interesting (and perhaps redeeming?) quality of this proposal is that it is proposing to give a standard meaning to

std::string str = "hello world";
std::vector<char> vec = std::move(str) | ranges::to<std::vector>();

From a human point of view, we can see that a quality implementation should implement that as a couple of pointer swaps, and ta-da, we have the “awesomer” buffer-ownership-transfer primitive described above! At least for transferring ownership between containers that are both provided by the same library (and possibly it has to be the same library that provides the implementation of ranges::to, I’m not sure).

However, Section 10 of P1206R1 implies that this facility will always copy — it is proposed to take its input range always by const lvalue reference — so, no move semantics for you! Also, the people who are most interested in speedily converting vectors to strings or vice versa are probably the same people who are least interested in pulling in all of <ranges> (with its 8-second compile time) just to get the facility.

For people who like their overload resolution faster than their ping time, note that P1206 proposes to be useable without operator| as well:

std::vector<char> vec = ranges::to<std::vector>(std::move(str));

Also, notice that the <std::vector> in the angle brackets is (A) not a typo and (B) nothing to do with CTAD. The idea is that the library implementor would write two different overloaded templates, both named to:

namespace nonstd {
    template<class Range>
    using range_value_t = iter_value_t<iterator_t<Range>>;
} // namespace nonstd

template<class ContainerType, class Range, class... Args>
to(Range, Args&&...);

template<template<class...> class ContainerTemplate, class Range, class... Args>
to(Range, Args&&...);

C++ continues to not-have (and likely will forever not-have) a way to encode “any arbitrary template with any arbitrary parameters.” So if you were hoping for

auto foo(std::span<int, 5> span)
    return span | ranges::to<std::array>;

then nope, that’s still Right Out.

Overall, I think P1206 probably should be accepted as part of Ranges. I like the paper’s “view materialization idiom.” I don’t like Ranges, or stomping on the name to, or clever metaprogramming shenanigans designed merely to further confuse the CTAD flock; but I think if we’re going to get Ranges then we should also get “view materialization” something like P1206.

Posted 2019-01-28