What's happening behind the scenes of Go's `nil` type? - go

Question

I recently read on a popular programming forum that Go supports a handful of "typelesss" values — notably, nil, Go's null value/bottom type. I have fairly limited experience with Go, and one of the statements made on this forum caught me off guard — namely, that it's illegal to write x := nil in Go.
Sure enough, a toy Go program with that line doesn't compile and the compiler error clearly points out that the comment from the forum checks out: declaration and assignment of a variable to nil is disallowed. This seems like a bit of an odd limitation, but things get stranger.
It's a common idiom in Go to return partial errors by returning a tuple from a function. Something like this:
func might_error() (int, error) {
return 1, nil
}
func main() int {
x, err := might_error()
if err != nil {
panic("err was not nil!")
}
return x
}
This has at least two inconsistencies as far as I can tell:
First, even though nil is typeless on paper, it takes on the error type (by way of it implementing the Error interface combined with Go's duck typing) for the purposes of conforming to might_error's return type signature.
Second, it seems like nil is being used for that previously-illegal declaraiton-and-assignment in main, and in a situation where (at least in comparable languages) might_error could be treated as a constexpr.
Weirder still, replacing x, err := might_error() with x, err := 1, nil still errors out with use of untyped nil!
My current line of thinking is that the Error interface is injected into a specific instance of nil in cases where a function's type signature requires it, meaning that it stops being an untyped nil and becomes a typed nil for the lifetime of that specific nil instance, but I'm not at all sure that this is correct since it seems like a strange design choice by nature of it not being clearly generalizeable.
What motivated these design choices? Why have nil be untyped instead of having it be a proper null type, except when it's convenient, at which point it becomes a typed value implementing an Error interface (to the best of my understanding)?

Answer 1

This is a complete non-problem in the sense that in practice there can be confusion if a nil value is stored in an interface variable making that variable non-nil. This is https://golang.org/doc/faq#nil_error and everyone is bitten a few times by this problem until you learn that interface values containing a nil variable are no longer nil themself. It's a bit like with var s = []*int{nil, nil, nil} where s contains only nils but is non-nil itself.
Technically (from a language design point) you could introduce several "nils", e.g. nil for pointers, null for interfaces, noop for functions and vac for channels. (Exaggerating a bit). With this you could have:
type T whatever
func (t *T) Error() string // make *T implement error
var err error // interface variable of type error
print(err == null) // true, null is Zero-Value of interface types
print(err == nil) // COMPILER ERROR, cannot compare interface to nil (which is for pointers only
var t *T = nil // a nil pointer to T
err = t // store nil *T in err
print(err == null) // false err is no longer null, contains t
You could even remove the compiler error:
err = t // store nil *T in err
print(err == null) // false, err contains t
print(err == nil) // "deep" comparison yielding true
err = &T{} // store non-nil *T in err
print(err == null) // still false
print(err == nil) // false, value inside err is no longer nil
You also could define a default type for nil, e.g. *[]chan*func()string so that you can write x := nil like you can do with f := 3.141 where the default type of the untyped constant 3.141 is float64. But such a default type for nil would be arbitrary and not helpful at all (as my example shows; *[]chan*func()string is uncommon).
If I remember correctly there was a longer discussion about this topic on the golang-nuts mailng list where the rationals about this design was discussed. It boiled down to something like: "The actual real-life problems with having nil multiple meanings and not being a constant are tiny (basically just variants of an error type containing nil not being nil). The 'solution' to this one tiny problem would complicate the language (e.g. by introducing a null literal for the zero value of interface types) considerably. It probably is simpler to teach people that interface values containing a nil are no longer nil themself than introducing typed nils or nulls for interface types."
In more than 10 years of programming in Go I literally never had to think about a nil literal being typed or untyped or constant or whatever. The article you probably are referring to is constructing pure academic, but actually non-problem in practice issue out of a tiny design decision about having just one nil literal for all zero values for pointer, slice, map, channel and function types.
Addendum
First, even though nil is typeless on paper, it takes on the error type (by way of it implementing the Error interface combined with Go's duck typing) for the purposes of conforming to might_error's return type signature.
This is a completely wrong description of what happens.
If you write
func f() (r int) { return 7 }
then 7 is assigned to r of type int and f returns. This works because 7 can be assigned to int.
In
func might_error() (int, error) { return 1, nil }
the same happens, the second return variable of type error (an interface type) is set to nil because nil can be assigned to any interface type like you can assign nil to any pointer type or any function type.
This has nothing to do with "implementing the Error interface combined with Go's duck typing". Absolutely not. nil doesn't implement the error interface at all. Any interface value can be nil like can be any function value or slice value or channel value. Setting a chan to nil basically "clears" the channel variable, it doesn't mean that nil somehow "implements the channel interface". You seem to conflate the zero value of several types and how to set them by assigning nil with implementing interfaces. All this has basically nothing to do with nil being typed or not. The nil literal in source code os overloaded and often can be thought of as just representing the zero value of a type.

Answer 2

Your example doesn't compile (because main needs to return nothing), and when you go into picky and nitty areas of a language, all your examples should actually work. 😀
My current line of thinking is that the Error interface is injected into a specific instance of nil in cases where a function's type signature requires it ...
That's not quite right. The Go specification tells us how this works:
There are three ways to return values from a function with a result type:
The return value or values may be explicitly listed in the "return" statement. Each expression must be single-valued and assignable to the corresponding element of the function's result type. [example snipped]
This is the method you use in your example program (which I cleaned up a bit to compile and run in the Go Playground). The expression:
return 1, nil
means:
unnamed_return_variable_1 = 1
unnamed_return_variable_2 = nil
where the two unnamed return variables don't actually have these names (they're un-named, after all) but do have types, courtesy of the declaration of the function. So
return 1, nil
is not at all like:
x, err := 1, nil
but rather more like:
var x int; var err error; x, err = 1, nil
which, as you can see, is also quite valid.
What motivated these design choices?
For that, you'd have to ask the designers.
ThinkGoodly mentions in a comment that:
The untyped 0 does not cause this angst.
The untyped constant 0 has a default type, though:
An untyped constant has a default type which is the type to which the constant is implicitly converted in contexts where a typed value is required, for instance, in a short variable declaration such as i := 0 where there is no explicit type. The default type of an untyped constant is bool, rune, int, float64, complex128 or string respectively, depending on whether it is a boolean, rune, integer, floating-point, complex, or string constant.
The predeclared identifier nil has no type, not even a default type. It just has a series of special rules that allow it to be used in various places.
The non-language-canon way to think about this
While this isn't how it's defined (it's defined by the Go specification), this is how I recommend thinking about the issue:
The predeclared identifier1 nil is untyped.
There are typed nil values. Anything sufficiently pointer-like can be nil, or can be non-nil. For instance, any *int variable can be pointer-to-int-nil, as an uninitialized one is. This nil is not the nil "keyword" (predeclared identifier, see footnote); it's a completely different "nil", just like Bruce (Banner) is a completely different person from Bruce (Wayne).
Interface values are two part: a type, and a value of that type. The type can be nil! This is yet another kind of nil, different from nil (the untyped "keyword") and nil (some particular nil of some particular pointer-y type). If the type is nil, the value can also be nil. If both are nil, the interface value compares equal to the <nil,nil> pair that the untyped "keyword" nil turns into when needed for comparison against an interface value. If either one is non-nil, the comparison says these are not equal. The hardware implementation is that the two parts of the interface are both zero: if either part is nonzero, the thing as a whole is nonzero, and hence "not nil".
The conversion from nil-the-"keyword" (predeclared identifier) to an appropriate hardware style zero value occurs when there's enough information supplied by context. Assignment to some variable—even an unnamed one—supplies context. Assignment to some positional argument supplies context. Trying to use a short declaration fails to supply context, hence the error.
1Go in general favors predeclared identifiers to keywords, so that false and true and so on are not actually keywords. This keeps nil out of the keyword set as well. Nonetheless, some things—such as the boolean constants false and true, or the magical behavior of the predeclared identifier iota, are only accessible by not covering up the predeclared identifier and then using it. That is, if you name something iota, you can't get the iota style functionality until that name goes out of scope. The way these things behave is often handled via keywords in other languages. If you tend to think in, say, C++ or Java, you might want to keep reminding yourself that these aren't keywords, even though they smell like them.