comex just noted that the design decision here has some tradeoffs. One of those tradeoffs is that you can't define wrapper types for units. Such a feature has clear use cases - it's not hard to find instances where unit conversion bugs have had severe consequences. "Most languages survived without that" is not a constructive comment.

Operator overloading is often misused and violates the principle of minimal surprise. You could even have side effects.

Still, type safety is totally possible. In Go, of your unit is Feet, you wrap your int into Feet and you cannot sum it with anything but feet. And that's all compile time, no cost.

Huge successful current languages don't have operator overloading now either. If anything, outside numerical/scientific work, it's even frowned upon.

And even in languages that do have them, it's not like people usually work with typed feet and such units. It is indeed safer, but I'm not sure it's mainstream (F#, Ada, Rebol, etc). Of course one can find such typed unit libs for other languages with Op. Overloading (C++ eg. has boost.units), but it's not like they're mainstream even where Operator Overloading is available.

In any case, it's not the first argument to make against a system language that it doesn't support Op. Overloading as a means for this use case.

In C++, if there weren't automatic destructor calls, you could still clean things up manually with a pseudo-destructor call, x.~T(), where T is the type of x.

In C, there are no (user-defined) generics, so you always know the concrete type of x, so you know whether it needs to have a destructor called and what the name of that destructor is.

In Zig, how would generic code create an instance of a generic type T that may or may not need to have a destructor called, ie. how do you know if you need to use defer and what the deferred expression should be? How would tooling be able to check that destructors were called if there is no special way of marking that a destructor exists?

I'm not sure how to reconcile this statement with the following claim from the linked article: "Zig has no macros and no metaprogramming yet still is powerful enough to express complex programs in a clear, non-repetitive way." Not only does it claim that Zig has no metaprogramming, it suggests that it sees metaprogramming as undesirable.

  fn max(comptime T: type, a: T, b: T) T {
      if (T == bool) {
          return a or b;
      } else if (a > b) {
          return a;
      } else {
          return b;
      }
  }

> You can always add a no-op dealloc method to your type if you want a generic function to handle it correctly

You're looking at it from the perspective of the generic consumer, not the generic author. The generic author generally does not have the ability to edit the type. Plus there are many types that do not have methods at all (integers, points, etc.).

> Moreover, Zig has awesome metaprogramming support so you can probably think of a way to check the type parameter for a dealloc method and only generate the dealloc call for those types.

Yes, you can detect and call a method called deinit that has the appropriate signature using @reflect. You can even put this in a function and call it a pseudo-destructor. But there's no guarantee that deinit has the actual semantics of a destructor without some kind of language-level agreement between class authors.

This will work fine. Do you have a more specific example?

To answer my question, no, they're not deinited. All deinit does is call self.allocator.free on the slice of elements, and for many allocators that's a nop. In fact none of ArrayLists methods take any kind of ownership of its elements. If you shrink an ArrayList(ArrayList(i32)) by one you leak the last ArrayList(i32). None of the methods call destructors on the elements because there is no generic notion of a destructor, only particular ad-hoc ones like deinit methods. ArrayList appears to solve the problem I mentioned above about generics not knowing if a generic type needs to have a destructor called by only supporting types that don't.

In C++, you'd write the function that clears a vector something like

    void clear() {
        for (auto p = ptr; p != ptr + len; ++p) {
            p->~T();
        }
        len = 0;
    }

I believe what he was wasking was "given an ArrayList(i32), how would you expect it to call deinit on the member i32s?". The answer, of course, is that you don't, which is also true of ArrayList(ArrayList(i32)). ArrayList absolutely supports heap-allocated types, you just have to free them yourself before calling deinit() on the ArrayList itself.

But if you did, you could use `@TypeInfo` to inspect the inner type for a function named `deinit`, or some other criteria that made sense for this determination.

I already mentioned that hack. All you're doing there is introducing the ad-hoc notion of a destructors as a method named deinit. Again, in order for that to work in generic code that convention needs to be blessed by the language.

Technically it only needs to be blessed by the standard library in this case.

C has had generics since C11. I’ve got a clone of “the good parts” of the std C++ template library (vector, unordered_map, ...) laying around somewhere.

    _Generic(x, type : value, ..., [default : value])

You usually wrap the call to '_Generic' in a macro. For instance:

    #define vec_push_back(vec, val) \
        _Generic(&vec, \
        vecint : vecint_push_back, \
        vecdbl : vecdbl_push_back, \
        default : assert_vwrap)(&vec, val)

The nice part, over C++, is that the generic instantiation is localized to a single translation-unit file, so you don't have to yell at the junior devs for instantiating templates without using an 'extern' construct.

For local variables, I'd say `defer` is an okay but not great substitute. If you call malloc() yourself, then sure, it's not hard to remember to stick a "defer { free(ptr); }" afterwards. But what if you call a function that returns an owned value? Can you tell just from looking at it that you're meant to free the thing afterward?

CoreFoundation, a C library on Apple platforms, has a cute solution to this: a strict naming convention. [1] If a function returns a value at "+1" reference count, i.e. you have to release it yourself, the function's name should include "Create" or "Copy", even where that would otherwise seem unnatural. For example, "CFSocketCopyAddress" returns a socket's address as a CFDataRef, returned at +1. You might expect it to be called "CFSocketGetAddress", but "Get" is reserved for functions that return values at +0.

As I said, cute – but fundamentally a workaround for language limitations. Distinguishing between two kinds of functions is exactly the sort of menial bookkeeping that computers excel at, whereas humans are prone to making mistakes. And – while CoreFoundation is forever stuck with the goal of C compatibility, Objective-C, which used to rely on a very similar system of manual retain/release and naming conventions, did ultimately transition to automatic reference counting.

But it doesn't have to be automatic. One use for destructors is to automatically release locks, and an interesting alternative design for that exists in Clang's "thread safety" annotations. These are annotations you can add to existing C or C++ code that uses mutexes; their main purpose is to catch you if you access a variable protected by a mutex without owning the mutex, but they can also complain if you lock something and forgot to unlock it. You can annotate functions as "requires mutex" (i.e. you must call it with the mutex locked, and it stays locked after the function), "acquire" (call with it unlocked, returns with it locked), "release" (the opposite), and so on. I think I once tried to abuse the annotation system to enforce reference counting, but it was broken. But I'd love to see a more thorough and robust design along these lines.

[1] https://developer.apple.com/library/archive/documentation/Co...