I don't disagree that Go's generics are pretty limited. But I find it a strange complaint, when contrasted with C++ templates. Which, as I understand, are literally not part of the type system and thus there seems to be a far stronger case, that they can not be called generics.

The main difference between Go's generics and C++ templates (and where some of the restrictions come from) is that Go insists that you can type-check both the body of a generic function and the call to it, without one having to know about the other. My understanding is, that with C++ templates (even including concepts), the type checking can only happen at the call-site, because you need to know the actual type arguments used, regardless of what the constraints might say.

And this decision leads to most of the complaints I've heard about C++ generics. The long compile times, the verbose error messages and the hard to debug type-errors.

So, if you prefer C++ templates, that's fair enough. But the limitations are there to address complaints many other people had about C++ templates specifically. And that seems a reasonable decision to me, as well.

> the type checking can only happen at the call-site, because you need to know the actual type arguments used, regardless of what the constraints might say.

No longer true after C++ 20. When you leverage C++20 concepts in templates, type-checking happens in the template body more precisely and earlier than with unconstrained templates.

In the below, a C++ 20+ compliant compiler tries to verify that T satisfies HasBar<T> during template argument substitution, before trying to instantiate the body

    template<typename T> 
      requires HasBar<T>
    void foo(T t) {
      t.bar();
    }

The error messages when you use concepts are also more precise and helpfully informative - like Rust generics

As I said in the other comment, I'm not a C++ user, so I'm relying on cargo-culting and copy-paste. But I think gcc disagrees - otherwise this would not compile, as line 14 is provably invalid: https://godbolt.org/z/P8sWKbEGP

Or am I grossly holding this wrong?

You need to have something that uses those templates. In your godbolt example, add a struct S

    struct S {
      bool M() { return true; }
    };


    int main() {
      S s;
      foo(s); // this now will check foo<S>
    }

Now you will get compile errors saying that the constraint is not satisfied and that there is no matching function for call to 'bar(S&)' at line 14.

> You need to have something that uses those templates.

Exactly. That is what I said:

> because you need to know the actual type arguments used, regardless of what the constraints might say.

It is because type-checking concept code is NP complete - it is trivial to check that a particular concrete type satisfies constraints, but you can not efficiently prove or disprove that all types which satisfy one constraint also satisfy another. Which you must do to type-check code like that (and give the user a helpful error message such as “this is fundamentally not satisfiable, your constraints are broken”).

And it’s one of the shortcomings of C++ templates that Go was consciously trying to avoid. Go’s generics are intentionally limited so you can only express constraints for which you can efficiently do such proofs.

I described the details a while back: https://blog.merovius.de/posts/2024-01-05_constraining_compl...

There are common solutions for the library issue. Authors of libraries for example can force instantiations for a dummy type that checks their concepts.

  template void foo(Dummy);

This can be done at the consumer side as well. I don't see a big deal of this. Dummy checks are common in Go too. For example, to check if a type satisfies an interface.

   var _ MyInterface = (*MyType)(nil)
   var _ SomeInterface = GenericType[ConcreteType]{}

After all, Go checks that a type implements an interface only at the point where you assign or use it as that interface type.

Thanks for your blog post. Unfortunately, the intentional limitations make the design space a massive headache and many times lead to very convoluted API. I would actually make the argument that it explodes complexity - for the developer, instead of constraining it.

> There are common solutions for the library issue. Authors of libraries for example can force instantiations for a dummy type that checks their concepts.

But that just ensures that the code type-checks for `Dummy`. It doesn't ensure that the code type-checks for any type you can put into `foo`. And that's the point of type constraints: To give you the necessary and sufficient conditions under which a generic function can be used.

That is simply not the case with C++ templates and concepts. That doesn't mean you can't still like them. I'm not trying to talk you out of liking C++ or even preferring it over Go. I'm just trying to explain that C++ concepts where something that we looked at specifically and found that it has properties that we don't like. And that - to us - the fact that Go generics are limited in comparison is a feature, not a bug.

And let's not forget that despite specifically reducing the safety of concepts in this way, the design ended up being NP-complete anyways and you can make a compiler use functionally infinite memory and time to compile a very small program: https://godbolt.org/z/crK89TW9G

For a language like Go, that prides itself on fast compilation times it is simply unacceptable to require a SAT solver to type-check. Again, doesn't mean one has to dislike C++. But one should be able to acknowledge that it is reasonable to choose a different tradeoff.

> I would actually make the argument that it explodes complexity - for the developer, instead of constraining it.

The title is a pun. Because it is about the computational complexity of checking constraints.

> But that just ensures that the code type-checks for `Dummy`. It doesn't ensure that the code type-checks for any type you can put into `foo`

Sure, that is C++ specific design decision. Just like Go made the design decision of not type checking interfaces leading to tens-of-thousands of lines of dummy checking concrete types against interfaces in popular Go repos.

I understand the design thinking even if I don't fully agree as a standard user of Go. Thanks for the detailed explanation in the blog.

Minor nitpick: It isn't all that difficult to come up with type structural/generic edge cases for ANY language compiler where compilation takes forever and times out in a playground. Here is a small program of ~100 lines leveraging Go Generics: https://go.dev/play/p/XttCbEhonXg

This will build for several minutes on your laptop if you use `go build`. It can be easily extended to several hours with a few modifications.

> Minor nitpick: It isn't all that difficult to come up with type structural/generic edge cases for ANY language compiler where compilation takes forever and times out in a playground. Here is a small program of ~100 lines leveraging Go Generics: https://go.dev/play/p/XttCbEhonXg

Fair point

Just to clarify why this is a problem: it’s possible for foo and bar to be defined in different libraries maintained by different people. Potentially several layers deep. And the author of the foo library tests their code and it compiles and all of their tests pass as and everything is great.

But it turns out that’s because they only ever tested it with types for which there is no conflict (obviously the conflicts can be more subtle than my example). And now a user instantiates it with a type that does trigger the conflict. And they get an error message, for code in a library they neither maintain nor even (directly) import. And they are expected to find that code and figure out why it breaks with this type to fix their build.

Or maybe someone changes one of the constraints deep down. In a way that seems backwards compatible to them. And they test everything and it all works fine. But then one of the users upgrades to a new version of the library which is considered compatible, but the build suddenly breaks.

These kind of situations are unacceptable to the Go project. We want to ensure that they categorically can’t happen. If your library code compiles, then the constraints are correct, full stop. As long as you don’t change your external API it doesn’t matter what your dependencies do - if your library builds, so will your users.

This doesn’t have to be important to you. But it is to the Go project and that seems valid too. And it explains a lot of the limitations we added.

C++ templates are duck typed at compile time.

Look at Haskell type classes or Rust's traits for some classic examples of how to 'type' your generics. (And compare to what Go and C++ are doing.)

> C++ templates are duck typed at compile time.

"Compile time" is not the right distinction. This is about "instantiation time". Go's implementation specifically allows to type-check the body and the call separately. That is, if you import a third-party package and call a generic function, all the type checker needs to look at to prove correctness is the signature of the function. It can ignore the body.

This is especially relevant, if you call a generic function from a generic function. For C++, proving that such a call is correct is, in general, NP-complete (it directly maps to the SAT problem, you need to prove that every solution to one arbitrary boolean formula satisfies a different boolean formula). So the designers made the conscious decision to just not do that, instead delaying that check to the point at which the concrete type used to instantiate the generic function is known (because checking that a specific assignment satisfies a boolean formula is trivial). But that also means that you have to (recursively) type-check a generic function again and again for every type argument provided, which can drive up compilation time.

A demonstration is this program, which makes gcc consume functionally infinite amount of memory and time: https://godbolt.org/z/crK89TW9G (clang is a little bit more clever, but can ultimately be defeated using a similar mechanism).

Avoiding these problems is a specific cause for a lot of the limitations with Go's generics.

> Look at Haskell type classes or Rust's traits for some classic examples of how to 'type' your generics. (And compare to what Go and C++ are doing.)

Yes, those are a different beasts altogether and the differences between what Go is doing and what Haskell and Rust are doing requires different explanations.

Though it's illustrative, because it turns out Rust also intentionally limited their generics implementation, to solve the kinds of performance problems Go is worried about. Specifically, Rust has the concept of "Dyn compatibility" (formerly "Object safety") which exists because otherwise Rusts goal of zero-cost abstractions would be broken. Haskell doesn't have this problem and will happily allow you to use the less efficient but more powerful types.

(All of this should have the caveat that I'm not an expert in or even a user of any of these languages. It's half-knowledge and I might be wrong or things might have changed since I last looked)

> C++ templates are duck typed at compile time.

This is not really true after C++ 20. C++ templates can leverage concepts that specify compile-time constraints and type checking on template parameters during template argument substitution

Yes, concepts are the fix to that duck typing. Not sure how widespread they are in the wild yet (nor how good compiler support is already at the moment.)