[burntsushi]

> But I know that fast compilers and runtimes that implement generics exist, so I don't find the whole discussion believable.

Are you arguing that an implementation of generics in a programming language is free? (Whether it be performance or implementation complexity.) Because if not, then you admit there are trade offs. If there are trade offs, then it is conceivable that some circumstances (including programmer tastes) may lead to a valuation that is different from yours.

"Performance" of the compiler is a critical goal of the Go project. Could you point out for me an industrial strength compiler that supports a type system of your liking that can match the compilation speed of the Go compiler? (I can't think of one.) Because if not, then I'll have to invoke this: talk is cheap, show me the code.

[tomp]

> Are you arguing that an implementation of generics in a programming language is free? (Whether it be performance or implementation complexity.)

I am.

Performance: if you replace all generics with interface{}, you get performance that is at better or equal to current Go code.

Implementation complexity: You could add some constraints to generics, for example only interface types can be used as generic parameters, and all uses of variables/methods/functions with generic types need to have generic parameters given. In that case, there's no complexity for code that doesn't use generics. Also, generics could actually be implemented as a preprocessor that would simply insert appropriate casts from/to interface{} at calls to generic methods. Since all variables/functions in Go are declared and have known types, you could of course go further and eliminate most generic parameter annotations.

[burntsushi]

> Performance: if you replace all generics with interface{}, you get performance that is at better or equal to current Go code.

That is incorrect. Using `interface{}` incurs a significant performance penalty at runtime because you end up boxing everything. For example, you cannot cast a `[]interface{}` to an arbitrary `[]A`.

In other words, no, `interface{}` is not `void*`, sorry. If it was, Go would not be memory safe. (See golang.org for their definition of memory safety. It is strictly superior to a language like C, but e.g., less than a language like Rust.)

[tomp]

What I meant is, "better or equal to current Go code using the current version of polymorphism (which is `interface {}`)".

AFAIK it's not possible to write generic Go code without `interface {}` (or, alternatively, duplicating the code for each type).

[burntsushi]

Yes, but the tradeoff Go takes right now is "slow programmers," or "no generics." You are trading that in for "slow programs." This directly implies that generics have some sort of cost structure and are not free.

`interface{}` is very rarely used for writing "generic" code precisely because of its lack of type safety and performance implications. If you came up with a way to make it type safe without addressing performance, you've just implemented a generics scheme that will cause programs to be slower than they would be without generics (because people would start using them!).

[tomp]

I think it should be clear to every programmer that there are different trade-offs to be made when designing systems, including the trade-off between generic-code&fast-development&slow-speed and specific-code&slow-development&fast-speed. People that don't get that shouldn't be programmers, and people that get it should have the freedom to make their own trade-offs.

In other words, the only "cost" of generics, thus, is the social cost (there is no purely technical cost). Here, it boils down to the philosophical/political freedom-vs-safety question, and I strongly incline towards freedom. Go does not. Go's position is, if not wrong, at least loosing, since programmers demonstrably are using work-arounds (`interface {}`) when they need generics.

[burntsushi]

> Go's position is, if not wrong, at least loosing, since programmers demonstrably are using work-arounds (`interface {}`) when they need generics.

I just got through telling you that this type of use is actively discouraged, so no, people aren't using it as often as you might think.

If you want generics, you've got to pay for it. The implementation you've suggested requires a runtime performance penalty to use generics. Therefore, you have not demonstrated that generics are free.

[gnuvince]

D's dmd supports templates and is faster than gc.

[burntsushi]

Is there any evidence for this claim? After a quick Google session, the best I could find was a forum thread 4 years ago claiming that `dmd` was faster than `gc`. Is that still true today?

[f2f]

Not anymore. You're two years out of date