Why does std::vector CTAD accept both int and double?

Question

#include <boost/type_index.hpp>
#include <iostream>
#include <vector>

int main()
{
    std::vector v{2, 3.14};
    std::cout << boost::typeindex::type_id_with_cvr<decltype(v)>().pretty_name() << '\n';

    std::cout << "size: " << v.size() << '\n';
    for (auto x : v)
        std::cout << "- " << x << '\n';
}

The output is:

std::vector<double, std::allocator<double> >
size: 2
- 2
- 3.14

std::vector instantiation uses auto-type-deduction (CTAD).

Apparently, the initializer_list constructor is used (checked), but why does it accept both an int and a double and still deduces the type as vector<double> ?
If I increase the number of elements (while mixing int and double), compilation fails.

The same result with gcc and clang

$ g++ --version
g++ (GCC) 13.0.1 20230401 (Red Hat 13.0.1-0)

$ clang++ --version
clang version 16.0.1 (Fedora 16.0.1-1.fc38)

@463035818_is_not_a_number Same result. size: 2, contains 5 & 3.14 — Tootsie, Apr 26 '23 at 14:26
What compiler are you using? What version of it? Have you tried with other compilers? What options do you pass when building, more specifically which C++ standard are you asking the compiler to use? — Some programmer dude, Apr 26 '23 at 14:26
@Tootsie right, I didnt pay attention, or I dont know what I did differerntly previously https://godbolt.org/z/7r8qTbdGe — 463035818_is_not_an_ai, Apr 26 '23 at 14:28
I would argue this is because a built-in ambiguity in the C++ specification with argument deduction and overload resolution. — Some programmer dude, Apr 26 '23 at 14:40
Check [cppinsights](https://cppinsights.io/s/a1c1a0ea) this uses constructor overload with initialization list! In brace initialization case initialization list has priority over other ctor overloads. — Marek R, Apr 26 '23 at 15:46

score 23 · Accepted Answer · answered Apr 26 '23 at 14:27

23

There are two instances of overload resolution happening here. The first is to deduce the template argument. Since the two types in the init-list are different, the size+value constructor (or more formally, the deduction guide generated from said constructor) is picked here, deducing the type to be double.

Then overload resolution runs again to create the object. With the type now known as double, the list constructor is now valid, and thus used to initialize the object.

answered Apr 26 '23 at 14:27

chris

60,560
13
143
205

1

So if you wanted to get a vector of size 2, filled with values all equal to 3.14, would you be forced to explicitly indicate the template argument as `std::vector`? – Fabio says Reinstate Monica Apr 26 '23 at 14:30
7

@FabiosaysReinstateMonica, I believe using parentheses instead of braces would do that because the list constructor is removed from the picture entirely. – chris Apr 26 '23 at 14:32
4

So you are saying that two different constructors are used for template type deduction and initialization. Shouldn't the same constructor be used for both actions? I can't get my head over this. – Karen Baghdasaryan Apr 26 '23 at 16:29
@KarenBaghdasaryan, It is a little weird. I can't say why it doesn't require that. – chris Apr 26 '23 at 16:43
Can you provide any section in standard with will address this issue, or at least some non-official resources? I would be very thankful. – Karen Baghdasaryan Apr 26 '23 at 20:00
2

@KarenBaghdasaryan, I was deliberately avoiding the standard itself on this one because there's no telling how long I'd be trapped in there, but I can offer [this CppCon video](https://youtu.be/-H-ut6j1BYU?t=1955) (at the relevant slide). I wonder if the proposal has motivation for allowing different results when a constructor or its generated guide is chosen in both cases or if the idea was to layer CTAD on top and that fell out as just a consequence. – chris Apr 26 '23 at 20:10

Why does std::vector CTAD accept both int and double?

1 Answers1