proposal: Go 2: revisit explicit annotations in :=

[jimmyfrasche]

Before Go 1 := would only work if every name on the lhs had not been declared in the current scope. This made it hard to write common code like

x, err := f()
//...
y, err := g()

Either y would require an explicit declaration with var or there would need to be an ugly workaround such as

y, err2 := g()
err = err2

Among other peccadilloes this meant copying a bit of valid code from one place to another could make it invalid and it would need to be rewritten.

This made := much less useful than it could be so the current behavior of implicitly reusing existing names defined in the current scope was introduced and erroring out if none of the names were new.

This made := much more useful, but has made it too easy to shadow variables in outer scopes.

When it was introduced, there was much discussion in the mailing list thread announcing the change (apologies but I cannot find a link) for a way to explicitly annotate the names to reuse such as

y, ^err := g() // mnemonic for "reusing variable defined above"

This was dismissed—if I recall correctly (again apologizes for not being able to track down the original thread)—because it would mean that copying a bit of valid code from one place to another could make it invalid, as annotations would need to be added or removed depending on the new context.

However, this is subtly true with the current behavior := depending on the names involved and what is or is not in the current scope, leading to bugs caused by accidental shadowing that are hard to track down due to the implicit nature of what is shadowed when. This is particularly vexing with named returns.

With explicit annotation it would still be possible to accidentally shadow a variable, but it would be easier to see what was happening at a glance and possibly easier to detect mechanically.

I think the idea of explicit annotation should be revisited for Go 2. I propose the following semantics

1. x, y, z := f() follows the pre-Go 1 semantics—all names must be unbound in the current scope
2. x, ^y, ^z := f() reuses y and z as defined in the current scope or an outer but non-global scope†
3. annotating a name that does not exist is an error
4. at least one name on the lhs must not be annotated

† making it incompatible with reusing variables in package global scope is a bit uneven but seems a wise safety precaution

This should be entirely go fix-able as the correct annotations to mirror the Go 1 semantics could be applied mechanically.

[bradfitz]

Dup of #377?

[jimmyfrasche]

That thread covers a lot of potential changes to :=. It looks like the explicit annotation was brought up after a dozen or so posts and eventually dominates the conversation (I only skimmed). I don't mind this being closed as a dup but #377 is closed due to age.

[davecheney]

Before Go 1 := would only work if every name on the lhs had not been

declared in the current scope. This made it hard to write common code like x, err := f() //... y, err := g() This is not correct, https://play.golang.org/p/aD8_f2r8M4

…

On Fri, Aug 4, 2017 at 10:04 AM, Brad Fitzpatrick ***@***.***> wrote: Dup of #377 <#377>? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#21303 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAAcA9ajSza1u0_AHhT0L619SCUtrY4qks5sUmAjgaJpZM4OtIK3> .

[jimmyfrasche]

@davecheney that's Go 1 code.

[Azareal]

^ is the bitwise xor operator.

I'm not really sure this is that useful to be honest.
If I think something's going to be shadowed, I usually explicitly declare the variables.

[bcmills]

Another alternative would be to annotate the new names instead of the old: that would make a ShortVarDecl equivalent to an Assignment.

For example, we could shift the colons from the = symbol to the new names:

x:, err: = f()
// …
y:, err = g()

[jimmyfrasche]

I don't really care about the syntax. There are other suggestions in #377 and the mailing list thread linked in there. I'm sure the Go team will pick something reasonable and pleasant in the face of all other constraints, if they decide to implement.

I'll use ^ in any examples that I write for consistency with the first post. I know it's an operator already but I think the polysemy should be fine in this case since a ^ would already be illegal in any place where this would come up, but, if it's not, that's fine.

I just want to be able to explicitly and concisely reuse variables from outer scopes without having to entirely and awkwardly rewrite the line with multiple var declarations. I want it to be easier to spot and avoid shadowing.

Annotating new names rather than old names would make it easier to avoid accidental shadowing in the first place.

I was thinking that another option would be to keep the default rules of := as it is now, but allow annotations to mean "use the one in an outer scope". That way Go 1 and Go 2 code would be source compatible (wrt this feature, at least) without needing a go fix. I'm not sure how far the plan is to go with that since that would mean no new keywords. It would mean you'd only be able to annotate old names, though.

I'd prefer "you always need to annotate" as it would mean not seeing an annotation had as much meaning as seeing an annotation, even though that would mean having to annotate more often.

[as]

I don't really care about the syntax.

Then why not just use

x, err := f()
y, e2 := g(); err=e2

[rogpeppe]

Before Go 1 := would only work if every name on the lhs had not been declared in the current scope.

I don't believe this is true, although I remember the discussion on changing the semantics of :=.
See my earlier comment: #21161 (comment)

From that previous discussion, ISTR favouring a syntax like this:

:a, :err = foo()     // equivalent to a, err := foo()
:b, err = foo()      // b is newly declared; err is not.

This doesn't avoid accidental shadowing though - I'm pretty sure you
don't want to ban all shadowed variable names.

[jimmyfrasche]

@as I don't care about the specific syntax used as long as there is some syntax with the desired semantics

@rogpeppe I could easily be misremembering or confusing it with a comment from the Go team that "it worked like that but then we changed it". Maybe even another language entirely. I tried to find the thread to check before posting but could not so I crossed my fingers and posted. Oh well.

At any rate, no I don't want to ban shadowing.

Shadowing is good. But it can also be awkward and easy to shadow too much.

The proposal is for some way to keep the nice bits of := but add a way to explicitly allow reuse of a variable defined in an outer scope. The current behavior can cause subtle problems especially with named returns and common names like err so use of := requires a dose of paranoia.

If that's the :a, err = foo() syntax, great.

[earthboundkid]

I would like it if shadowing an outer scope in the short declaration were banned.

So, for example, this would be legal:

func Example() (int, error) {
    x, err := thingOne()
    if err != nil {
        return 0, err
    }
    y, err := thingTwo()
    if err != nil {
        return 0, err
    }
    return x+y, nil
}

Because while err is reused, it does not shadow an outer err.

But this would be illegal:

func Example() (int, error) {
    x, err := thingOne()
    if err != nil {
        return 0, err
    }
    if x > 1 {
        y, err := thingTwo()
        if err != nil {
            return 0, err
        }
        x += y
    }
    return x, nil
}

Because the inner err shadows an outer one.

I think something like a shadow ban is necessary to prevent the footgun of accidental shadowing.