Why does putting concept to type specifier fail type deduction when the same constraint must deduce different types?

Question

We have:

template <typename ...T> concept bool Numerics = ( std::is_arithmetic_v<T> && ... ) ;
template <typename T>    concept bool Numeric  =   std::is_arithmetic_v<T>;

Then we let compiler deduce all the numbers:

template <typename T, typename U, typename V, typename W> requires Numerics<T,U,V,W>
auto foo(T arg1, U arg2, V arg3, W arg4) {
    return 0.0 + arg1 + arg2 + arg3 + arg4;
}

std::cout << foo (1,2.0,3.0f,4.0l) << "\n";

Compiler deduces all argument types like expected:

auto foo<int, double, float, long double>(int, double, float, long double):

When we try distribute constraints into type specifier for writing a shorter version:

auto foo2(Numeric arg1, Numeric arg2, Numeric arg3, Numeric arg4) {
    return 0.0 + arg1 + arg2 + arg3 + arg4;
}

Though, compiler surprisingly fails to deduce this:

// err: deduction fails
//
// std::cout << foo2(1,2,3,4.0) << "\n"; 

It seems compiler try deduce all into the same type, which must fails here.

Why? Shouldn't compiler be able to deduce different types from the same contraint?

LIVE

Answer 1

This is one of the more controversial aspects of the Concepts TS. When you say

template <typename T> concept bool C = true;

template <C T, C U>
void func(T t, U u);

this is translated as if you had said

template <typename T, typename U> requires (C<T> && C<U>)
void func(T t, U u);

However, when you use the "abbreviated syntax" to say

void func(C t, C u); 

this is translated as if you had said

template <typename T> requires C<T>
void func(T t, T u);

Many people consider this to be counterintuitive, and there has been a proposal to change it. However, others (including Bjarne Stroustrup himself) are in favour of "consistent deduction" -- see Stroustrup's paper here.

At the time of writing, the subset of the Concepts TS incorporated into the C++20 draft does not include the abbreviated syntax. It's not clear at the moment whether it will end up in '20, and if so, what sort of deduction it will use -- the arguments are still to be resolved.