Go-internals: Chapter 2, “Interfaces” released

[ivoras]

It surprised me how go's interfaces have non-trivial limitations. What lately bit me is illustrated in the following snippet: the call to doSomething() is just not possible in Go:

  package main

  import "fmt"

  type S struct {
        i int
  }

  func (s *S) foo() {
        s.i++
        fmt.Println(s.i)
  }

  type Sarr []S

  type I interface {
        foo()
  }

  func doSomething(arg []I) {
        arg[0].foo()
  }

  func main() {
        arr := Sarr{S{i: 1}, S{i: 2}, S{i: 3}}
        arr[1].foo()

        doSomething(arr)
  }

I understand why the current implementation of Go doesn't work that way, and that the meta-reason for this not being implemented is the "we don't want generics" attitude, since it would effectively require 2 different compilations of doSomething(), one which accepts a slice of interfaces, which are a (pointer,type) tuple, and one which accepts an slice of this specific struct which happens to actually implement the I interface, and that's a taboo, but come on... it just weakens the idea of interfaces. It could have been done as a compile-time work-around which constructs the slice-of-interfaces argument on-the-fly before calling the function - it wouldn't have been the only magical thing in Go.

[leaveyou]

>It could have been done as a compile-time work-around which constructs the slice-of-interfaces argument on-the-fly before calling the function - it wouldn't have been the only magical thing in Go.

Is there a language that works the way you describe ?

[perfmode]

The reason this doesn’t work is because conversions of this nature may require allocations, and like C++, Go doesn’t want to hide costly operations from the author.

[paraboul]

Really nice write up.

I wish more microbenchmarks go that deep into IR/codegen analysis (somewhat related : https://mrale.ph/blog/2014/02/23/the-black-cat-of-microbench...).

[jnordwick]

Go assembly is more of a low-level intermediate representation. It isn't as high level as LLVM-IR, but it isn't what the processor runs on. I wonder how much is obfuscated by this, especially when compiled with optimizations?

Is there really much to be learned by analyzing performance in unoptimized pseudo-assembly? It's like looking at JVM bytecode for performance indicators.

[pansinghkoder]

Love the writeup! Very readable for someone like me who works much layers up in programming.

I have also written a primer on go-concurrency. https://gist.github.com/rushilgupta/228dfdf379121cb9426d5e90...

Thoughts?

[timtosi]

Looking forward to read the GC chapter.

[kierkeGuard]

Great writeup, really happy to see closer looks at Golang. I've worked with Go a few times, but I've thought that getting good independent looks/great examples is a bit too difficult.

Thanks!

[0xdeadbeefbabe]

What does //go:noinline mean in this context?

[axaxs]

go has certain quasi hidden directives for the compiler. This one means "don't inline the following function."

[sAbakumoff]

to me it seems like the version of "you don't know JS" but for Golang. nicely done!