> I would class this more as 'duck-typed templates'.
Or parametrically polymorphic type variables!
> This is an orthogonal issue, to do with whether instantiations of parameterized types/functions are monomorphised or not.
I stand corrected, as monomorphization is optional for parametric polymorphism but mandatory for ad hoc polymorphism. It is still technically possible to deduplicate similar enough functions, either at the source level or at the binary level, but that is orthogonal to this issue :-)
> I would say the big issue with templates is that they are slow and produce hard-to decipher compile-time stack traces, because they are not checked at the instantiation site.
You are right for the status quo of C++ templates, but I'm not sure for general cases. It might be possible to keep error information down to the instantiation site for better error messages... or not.
> I stand corrected, as monomorphization is optional for parametric polymorphism but mandatory for ad hoc polymorphism.
It's not mandatory for ad-hoc polymorphism. Haskell desugars ah-hoc polymorphism to dictionary passing (although it often does inlining to reduce the overhead of this).
>I stand corrected, as monomorphization is optional for parametric polymorphism but mandatory for ad hoc polymorphism. It is still technically possible to deduplicate similar enough functions, either at the source level or at the binary level, but that is orthogonal to this issue :-)
Regarding this bloat thing, are monomorphized functions ever generated for types/signatures that are not actually _used_ in the program?
If not, it's not really bloat, is it? It's what one would do themselves without generics, if they wanted to be able to use them with full speed.
> It's what one would do themselves without generics, if they wanted to be able to use them with full speed.
Right, but in many cases they can be combined without much speed penalty. For example you can combine monomorphized functions when you only rely on methods that are expected to be costly and inlining has a negligible effect. With actual, not ad hoc, generics you may (hopefully) have a control over monomorphization---I think Rust's trait object is a good design for this reason.
It is certainly possible. Concept C++ (the failed C++0x concept proposal) did it; so, to a certain extent, does concept 2.0. Rust which was designed from day 1 to support this also does.
Or parametrically polymorphic type variables!
> This is an orthogonal issue, to do with whether instantiations of parameterized types/functions are monomorphised or not.
I stand corrected, as monomorphization is optional for parametric polymorphism but mandatory for ad hoc polymorphism. It is still technically possible to deduplicate similar enough functions, either at the source level or at the binary level, but that is orthogonal to this issue :-)
> I would say the big issue with templates is that they are slow and produce hard-to decipher compile-time stack traces, because they are not checked at the instantiation site.
You are right for the status quo of C++ templates, but I'm not sure for general cases. It might be possible to keep error information down to the instantiation site for better error messages... or not.