The center for all discussion and news regarding C++.
It's an interesting little gotcha, but I wonder if it wouldn't be preferable to just implement a type trait like is_expensive_conversion that is false by default and true for specific cases, and just static_assert this stuff.
Relying on obscure implicit type conversion mechanics for this feature feels like being too clever for your own good.
That's a fair criticism around relying on implicit type conversion mechanics, and part of the tradeoff to make. On the other hand, I imagine (and my imagination may be limited) that one downside of static_assert is to increase verbosity, something like:
auto r = f();
static_assert(std::is_same_v<decltype(r),MyReturnType>> || !is_expensive_conversion_v<MyReturnType>);
return r;
I'm wondering where you wanted to use this. Was it a generic context, where you didn't know what types you were getting? Also, I think it will interfere with the retern-value optimization.
Yes, that's right, generic context, and you may be right on return value optimization. It was for implementing a collection of numerical functions that take array arguments, where the elements of those arrays could be of various arithmetic types, and the return type should be an array of a particular arithmetic type given promotion etc. The implementation was generic, and I was wanting to validate its correctness wrt return values having the correct arithmetic type without implicit copy.
For the array type it can be useful to allow implicit copy to different arithmetic types (design choice, I'm now back to explicit constructors to disallow this for what it's worth). If allowed though, I still wanted a compile time check like this to ensure that it wasn't happening by accident in particular circumstances.