Hacker News Reader

Comment by Alupis | Hacker News Reader

I have to admit, the STL - and more generally templates - in C++ are very perplexing to someone that has a strong background in a language with a more complete generics implementation (ie, Java or C#).

The very fact that templates are not classes, specifically not base classes for all derived variants of the template is perplexing and, in my opinion, a major swing and miss for a language that seems to pride itself on having everything including the kitchen sink.

I've read the reasons behind why templates do not share any inheritance with derived classes, and it makes sense within the context of how C++ and more specifically the compiler works... but it limits the usefulness of templates past mere containers.

View on HN

> but it limits the usefulness of templates past mere containers.

Your post (specifically, your conclusion) makes me think that you are fundamentally mistaken about how to use templates.

Template metaprogramming is an old-trick at this point, with a huge variety of applications in C++. But even if you ignore Template metaprogramming (which many argue is too difficult to understand...), bog-standard C++ techniques like CRTP (Curiously Recurring Template Pattern) shows how useful templates can be outside of containers.

https://en.wikipedia.org/wiki/Curiously_recurring_template_p...

---------

C++ templates are extremely different beasts than generics. Templates are closer to a meta-language than a language feature. They hold important roles in optimization... specifically allowing you to convert "runtime code" into "compile time code" in a huge class of problems.

CRTP is a great example of retaining OOP-style polymorphism while compiling down into highly efficient non-virtual calls... with practical applications into GUI frameworks, video games and more. And its only made possible due to templates.

Interesting, I've never heard of CRTP and will need to read up on it.

I do "get it" that templates are not generics... there's just a lot of gotchas for someone assuming generics, err - templates are similar to other languages. Like the difficulty in nesting templates in templates...

Even in the use of "containers", generics are the complete opposite of how to use templates.

Java / C#'s concept of "Everything derives from Object" is outright rejected in C++. Templates are part of this culture: how can we provide extendibility when nothing is an Object?

Well, look at iterators in C++. All C++ iterators extend from all collections WITHOUT the use of OOP-derivations. That's thanks to the magic of Templates.

------

What's going on? C++ templates allow the creation of new iterator classes. It invokes the compile-time language to create a new class at compile time.

In Java / C#: vector<MyClass> is the same old vector as all other vectors.

In C++: vector<MyClass> is a custom-made, completely separate class from all other vectors. vector<MyClass>::iterator is a 2nd, custom-made completely separate class from all other iterators.

And yet, all of these different classes can be used the same, thanks to the use of auto, or other compile-time structures in C++.

Getting Java's generics out of your head when doing C++ is definitely a challenge. It's just so disappointing you cannot do something like:

    void myFunc(List<? extends MyDataContainer<?>> list);

Instead you end up with a base class MyDataContiner and then manually extending it with something like MyDataContinerInt or MyDataContinerLong etc...

C++ templates may be very powerful, but limitations like this feel out of place in a language as bloated as C++ already is...

Admittedly, I still have much to learn about C++. It's been quite an experience, frustrating at times, learning this language.

I think you can get something similar now with C++20 concepts?

    template<typename T, typename U>
    void my_func(std::vector<T>* vec) requires std::derived_from<T, my_data_container<U>>;

(Probably mangled something, but the idea is there)

Without concepts, you're left with SINFAE:

    template<typename T, typename U, typename = std::enable_if_t<std::is_base_of_v<my_data_container<U>, T>>>
    void my_func(std::vector<T>* vec);

or just a regular function and a hope that anything wrong will be caught by the compiler:

    template<typename T, typename U>
    void my_func(std::vector<T>* vec) {
        // use my_data_container<U> somewhere and hope things blow up if something is wrong
        // Not 100% sure this is allowed, but something similar should be at least
    }

I'm curious: what's wrong with myFunc(List<BaseClass * >) ??

EDIT: Remember, in C++, "BaseClass" cannot polymorph, only "BaseClass * " can. (BaseClass is where the data is stored precisely. If its 20 bytes of RAM, then DerivedClass might be 24 bytes of RAM, and therefore can't fit. But both BaseClass * and DerivedClass * are compatible with each other, and may convert into each other, as long as you know the pointer goes to the right kind of class)

And if that's not sufficient, then a static_assert over the type information is probably what you need. Ex:

   template<class T>
   void myFunc(List<T> list){
      static_assert(std::is_nothrow_convertible(T, BaseClass)); // I haven't tried, but this probably works
   }

That's what I mean about C++ templates: they're a meta-language. You can add if-statements / else statements over type information at compile-time in C++.

In this case, we've created a static_assert (aka: a compile-time error will pop up) if T does not convert into BaseClass (note: T may have a CopyConstructor(BaseClass), so maybe its not exactly tthe same concept as what you're trying to do... but this gives you an idea of C++isms...)

> I'm curious: what's wrong with myFunc(List<BaseClass * >) ??

Well, nothing I suppose, except BaseClass cannot then be a template, and you cannot have derived class-specific return types like a template would allow.

So, having BaseClassInt type return int from some function but BaseClassLong return a long gets more challenging than it should be. (I recognize all my examples so far had a void return type, that was just for simplicity).

> But both BaseClass * and DerivedClass * are compatible with each other, and may convert into each other, as long as you know the pointer goes to the right kind of class)

Which loses type safety, unfortunately.

> In this case, we've created a static_assert (aka: a compile-time error will pop up) if T does not convert into BaseClass

I'll look into the static_assert() option, could prove useful in some cases.

At the end of the day, there are two things at play here - my lack of C++ familiarity, and in some cases it's just how C++ is.

It is a strange feeling though, having a language limit your expressiveness for the first time...

> Well, nothing I suppose, except BaseClass cannot then be a template

    template <class T>
    void myFunc(List<T*>){
        static_assert(stuff about T that you want to guarantee at compile time); 
    }

See here: https://en.cppreference.com/w/cpp/header/type_traits

For a list of common functions used for compile-time type checking.

Thank you for this conversation. You've given me some homework to read up on, and I appreciate it a lot!

Templates are the killer feature of C++. They add incredible power to the language, but also dramatically increase its complexity. Templates are only superficially similar to generics in C#/Java. They are essentially a type-safe, Turing-complete, syntactically-constrained macro system deeply embedded into the grammar of the language.

Few other programming languages offer a similar feature. C++ pulled it off with an ISO standard and multiple conformant implementations.

Templates are loved by many, but they are a contender (along with UB) for the most hated C++ feature due to complexity.

As a C++ developer, I see things entirely the other way. Java templates are much less powerful than C++ templates. In Java, templates were basically just a convenient way to do type erasure. That's better than nothing, but the inability to specialize templates makes them quite limited. In C++, you can define a common base class if you need one, but being able to optimize based on the template type is very useful.

Function templates are great for writing concise, efficient math that handles half-precision, single-precision, double-precision, single-precision complex numbers and double-precision complex numbers (and more) with minimal duplication.

Templates are also what make interfaces like std::format efficient and easy to use.

> The very fact that templates are not classes, specifically not base classes for all derived variants of the template is perplexing

I'm a little confused at what you're trying to get at here. What do you want to do that needs such a thing? What is the equivalent Java/C# code you have in mind?

> but it limits the usefulness of templates past mere containers.

What kinds of uses did you have in mind for a "template base class"?

Extend a template and specialize it for some type. Extend the template again and specialize it for some other type.

The two resulting derived classes are not the same, and you cannot have a function that accepts the base template with a wild card type and pass both types into it.

In Java, you can do: void myFunc(List<?> list); and all List objects with any type can be accepted by this method, in addition any class that extended List can also be accepted into this method, such as ArrayList<?> or whatever.

As far as I've seen, this is impossible in C++, unfortunately.

Why not

    template<typename T>
    void my_func(std::vector<T>& vec);

The reference (or use a pointer, if you want) ensures that derived types can be passed, and the template allows for any parameterization to be used.

Reply on news.ycombinator.com