> This is a very interesting situation (and not a very common one for me) because it is where Go’s simple errors as values decisions shines. Doing this kind of thing with exceptions would be pretty clumsy, [...]
Just because a language supports exceptions does not mean you always have to use them. You can always use a helper function to catch any error and return it as a value.
Wouldn't that basically be replicating a return value-error approach, in a more cumbersome way though?
You could argue that Go can use Exceptions in the same manner as X lang too, by using panic and recover... but most people seem to not like that.
With Go and Rust, return value error handling is just simply nice. There's no question about if a function might error out, just look at the signature.
> Wouldn't that basically be replicating a return value-error approach, in a more cumbersome way though?
This would not be "replicating" anything, exceptions already are values in most languages. The only thing to do is wrap your call with a function, like "err = catch-all(expr)".
> ou could argue that Go can use Exceptions in the same manner as X lang too, by using panic and recover...
Except that it would be cumbersome ;-) because panic/recover works only in combination with defer[0].
> With Go and Rust, return value error handling is just simply nice.
Using non-exceptional control flow for errors can be useful depending on the circumstances. Using exceptions is simply nice in many other cases.
> There's no question about if a function might error out, just look at the signature.
If your language does not provide a similar tool, you can't just use a function because "expr" is evaluated before the function is called. I had Lisp macros in mind when writing this. A poor man's approach that is still generic is to wrap the expression in a closure:
I got bit by this the other day when returning a nil pointer to a struct as an error. It was incredibly frustrating when I finally figured out what was going on, and I'm curious if there's a good reason to distinguish between these 2 types of nil or if it was just an oversight/mistake during the language design process.
Go has the same complexity. From the Go spec: "For an expression x of interface type and a type T, the primary expression x.(T) asserts that x is not nil and that the value stored in x is of type T.'"
I had run into the same situation of handling multiple errors. Assuming I wasn't the first to need such a thing I found multierror[1]. Searching now returns more choices some of which use a nil []error and some with structs.
I love the idea of a "read only" map (or any other type), but it seems it is just a side-effect landmine in this case. Yuck...almost strikes me as a bug as described in the article
A map is a pointer to the actual map implementation. When declared via
var m map[T]T
you're basically writing
var m *runtime.hmap
which is a nil pointer. Go has sane zero values, so it makes sense that reading from a nil map returns the zero value for whatever type its values are.
However, there's a trade-off: either Go's maps allow nil assignment (in a similar fashion to slice's `append') _or_ it retains its "reference" (forgive me for using that term) semantics.
Go's designers decided that maps should differ from slices in that regard, and I tend to agree. I think it'd be a mistake to have to always return a map _just_ in it's newly allocated.
Except it's not a language implementation detail, it's part of what the language is. You might as well ask why a user of another language needs to know what a class is, or how inheritance works. Too many people seem to be assuming "Oh, it's like void*, so it must act like that always!", but it's not.
After using Go for a few years, I've learned (the hard way) that Go does not separate the concepts of implementation details and the official language specification. The official implementation is the language, and they're inseparable. This is why a lot of design decisions about the language make you say "wait, what? why does it need to be that way?" and the community's official answer is "because that's how it works best with how it's implemented".
Yes, but the whole point to using abstractions (such as high-level languages) is to separate the context where implementation details are relevant (e.g., when writing a compiler), from the rest of the system, where implementation details don't matter (e.g., when using a compiler).
It's a stunningly difficult and surprisingly low-value proposition to create abstractions that don't leak. That's not the goal anyway: The goal is to leverage abstractions to build things. The implementation details frequently do matter. Even if your abstraction is airtight, it has to run on a real machine. When possible, it's preferable to design abstractions that have a clear implementation in terms of the lower level's abstractions. This way, when something goes wrong - and it will - you can understand and fix it.
The behavior of Go's nil does not make the language any simpler to implement in any meaningful sense. The only thing it affects at the machine level is which cmp instructions are generated.
I won't defend the weird behavior of equality comparisons to nil, that definitely seems wrong. I was mostly just addressing the notion that a language needs a formal specification and must fully abstract the machine to be useful or at least not worthy of ridicule.
> It's a stunningly difficult and surprisingly low-value proposition to create abstractions that don't leak.
Maybe for you. For me, non-leaky abstractions have enormous value. And, while designing non-leaky abstractions requires more effort upfront, in the long run, it requires less effort than plugging leaks in carelessly designed (non)abstractions.
> That's not the goal anyway: The goal is to leverage abstractions to build things.
The goal is to do it efficiently, and creating as few problems as possible for the future.
> The implementation details frequently do matter.
Of course they do matter - for the implementor.
> When possible, it's preferable to design abstractions that have a clear implementation in terms of the lower level's abstractions.
No disagreement here.
> This way, when something goes wrong - and it will - you can understand and fix it.
That's the job of the abstraction's implementor. If that happens to be me, I will fix it. Otherwise I'll just submit a bug report. If that doesn't work, then I'll reimplement the abstraction myself, but by no means will I work around an existing broken implementation.
The last bit of information you need to understand that is that unlike some other languages you may be used to, "nil" is a perfectly valid value for a pointer to have. You can write methods like this:
func (o *Object) ReturnSomething() int {
if o == nil {
return 0
}
return o.SomeOtherValue
}
It is also therefore legal to return in an interface value a nil pointer to a struct of a particular kind, and therefore "an interface containing a nil pointer of a particular type" and "an interface containing nothing at all" are fundamentally different things that can not be collapsed together.
So as Vendan says, it is just part of the language. All languages have this sort of wart in them, where two or more perfectly sensible decisions interact to create something that isn't sensible at first glance.
(I'm still in favor of going back in time and changing Go to have non-nillable values, but that ship has certainly sailed.)
Java and C# also have interfaces and they don't have this strange behavior around null. I don't see how it's any sort of natural consequence of having null plus interfaces. Rather I suspect the reason is that the type equality operator was implemented to check one word and not the other when comparing two-word interface types, and the behavior that resulted was eventually specified instead of changing the implementation to compare both words.
> "nil" is a perfectly valid value for a pointer to have.
So, just like C and Java. I don't see the difference.
> You can write methods like this: (snippet)
Oh, now I get it: `nil` itself isn't a value. `nil` is syntactic sugar that expands into a nil value. (In other words, confusingly enough, `nil` and “nil value” are different things!) This is unlike Java, where `null` itself is a value.
It would be very helpful if people described things precisely.
> All languages have this sort of wart in them, where two or more perfectly sensible decisions interact to create something that isn't sensible at first glance.
This means that at least one of the two (or more) decisions was less sensible than it originally seemed. Good design decisions don't introduce warts into a system.
nil doesn't "expand", it's just what it is. You have to bear in mind that methods are roughly syntactic sugar from
func (o *Object) ReturnSomething() int
to
func ReturnSomething(o *Object) int
In the first case, it may "seem" that a nil Object would be invalid, but in the second case it's obviously valid, and makes complete sense. It may be described as a "wart", but it makes complete sense to me. The issue is, in my opinion, that people drag their own expectations into programming languages as they learn them.
In this case, Java and C# have the same feature and don't have this behavior around null. I think it's reasonable for people who have seen identical features in other languages to expect similar behavior when learning new ones.
They don't have the same feature of "I can call a method on a null value", at least, not that I've seen... You are calling them identical features when the fact is that they aren't. C# and Java both have generics, do they perform identical? Haskell has generics too, but I bet they'd throw a shit-fit if you tried to say "Oh, C# has generics too, so why would I bother with Haskell?"
The ability to call a method on a nil receiver doesn't have anything to do with what comparisons should do. Semantically, the question is whether == should compare the values or the vtables. In no other language I know of does it compare vtables.
If x.Foo() and y.Foo() do different things and return different values, clearly x != y, and golang needs some kind of SQL-style is-null check rather than using equality to check for null.
But dynamic dispatch on "what type of object do I not have?" is still a damn weird thing to want.
Interface comparisons check both, the value and the dynamic type. Describing that as "comparing vtables" is misleading.
It the types don't match, the operands are not considered equal. Type and value have to match for two interface values to be equal. How does that not make sense?
For mismatched dynamic types you can't meaningfully compare the values in the general case, so of course you need to consider types.
Having dealt with languages that don't, I much prefer ones that do. The thing is, most languages DO basically check vtables, in that 2 values with different types will never be equal (unless there is implicit casting or some kind of equality override). For instance, 1 == "1" is valid, and true, in PHP, even though that is less then ideal. The difference, of course, is that most of those checks are done at compile time in many languages. Go will do some checks at compile time, but it's specifically unable to do some of them due to the "dynamics" of how interface variables work.
yes, and it's a completely different behavior then Go, where it's perfectly defined as well: you call that method with a nil pointer as the receiver. Please don't act like those 2 behaviors are anywhere near similar.
An interface being equal to nil just because the value it contains is nil would be even more confusing. For example, imagine if you had two interface values, one holding a nil * T and one holding a nil * S. They would both be equal to nil, but not equal to each other, losing transitivity of equality.
Also, sometimes nil is a perfectly valid value of something and can implement interfaces. You can call methods on nil receivers because method calls are just curried function applications over the value. It would be odd if the interface compared equal to nil just because I happened to implement it with a value that's valid with nil.
Couldn't they also be equal to each other? If equality was defined only in terms of the value and not the vtable (to borrow pcwalton's terms), they would be equal. Are there any problems which would arise from this?
Equality is not necessarily defined for values of different types ("with different vtables", if you will). You can't just compare values of different type and expect a meaningful result.
In this case, though, we're talking about interfaces; an interface is a pointer plus this bit we're calling a vtable. Pointers, even to different types, absolutely can be compared for equality, can't they? More specifically, i am pretty sure pointers can be compared to see if they are both nil, which is all that would be needed here.
Sometimes you try and make one feature work a certain way and that decision has unintended consequences. Most languages have surprises like this.
In this case, it's an implicit conversion from "nil" to the "nil value" of a given type. This is convenient in many cases, but causes confusion in other cases.
It's also not an implicit conversion from nil to something. Nil isn't a value the way 0 is a value of an integer or "foo" is a value of string.
Per Go spec (https://golang.org/ref/spec) nil is "predeclared identifier which has no type". That implies it doesn't have a value and therefore cannot be converted to a value.
Nil is zero value of pointer types (which in practice end up being a pointer whose value interpreted as integer is 0) and also a zero value of interface type (which is interface without a type and without a value) and a few other types.
The worse is not, that. I think the worse is that, unlike Java, a race condition in Go has a totally undefined behavior that can corrupt the memory of the entire system your Go app is running on. You can effectively shut down your computer with a race condition in a Go program, and it actually happened to me many times on Windows, with 3rd party Go tools that didn't handle concurrency carefully. Go concurrency is truly unsafe, C style.
This shouldn't happen on any kernel with memory protection. Userspace corruption, sure, but shutting down the computer is unrealistic unless you are running Go code on Windows 98 or classic Mac OS.
A man orders coffee in a restaurant.
"Would you like it with or without cream?"
"Without"
A few minutes later the waiter returns.
"I'm sorry sir, we've run out of cream. Would you like it without milk instead?"