proposal: spec: reduce noise in return statements that contain mostly zero values

[jimmyfrasche]

Update: the current proposal is to permit return ..., v to return the zero value for all but the last result, and to return v as the last result. The most common use will be return ..., err.

A variant under discussion include return ..., v1, v2 to return zeroes for all but the last N.

Another variant is to permit return ... to return zeroes for all.

In general ... is only permitted if it omits one or more values--func F() err { return ..., errors.New("") } is not permitted.

---

Proposal

In return statements, allow ... to signify, roughly, "and everything else is the zero value". It can replace one or more zero values.

This is best described by example:

Given the function signature func() (int, string, *T, Struct, error):

return 0, "", nil, Struct{}, nil may be written return ...

return 0, "", nil, Struct{}, err may be written return ..., err

return 0, "", nil, Struct{X: Y}, nil may be written return ..., Struct{X: Y}, nil

return 1, "", nil, Struct{}, nil may be written return 1, ...

return 1, "a", nil, Struct{}, nil may be written return 1, "a", ...

return 1, "", nil, Struct{}, err may be written return 1, ..., err

return 1, "a", nil, Struct{X: Y}, err may be written return 1, "a", ..., Struct{X: Y}, err

The following is invalid:

return ..., Struct{X: Y}, ... — there can be at most one ... in a return statement

Rationale

It is common for a function with multiple return values to return only one non-zero result when returning early due to errors.

This creates several annoyances of varying degrees.

When writing the code one or more zero values must be manually specified. This is at best a minor annoyance and not worth a language change.

Editing the code after changing the type of, removing one of, or adding another return value is quite annoying but the compiler is fast enough and helpful enough to largely mitigate this.

For both of the above external tooling can help: https://github.com/sqs/goreturns

However, the unsolved problem and motivation for the proposal is that it is quite annoying to read code like this. When reading return 0, "", nil, Struct{}, err unnecessary time is spent pattern matching the majority of the return values with the various zero value forms. The only signal, err, is pushed off to the side. The same intent is coded more explicitly and more directly with return ..., err. Additionally, the previous two minor annoyances go away with this more explicit form.

History

This is a generalized version of a suggestion made by @nigeltao in #19642 (comment) where #19642 was a proposal to allow a single token, _, to be sugar for the zero value of any type.

I revived the notion in #21161 (comment) where #21161 is the currently latest proposal to simplify the if err != nil { return err } boilerplate.

Discussion

This can be handled entirely with the naked return, but that has greater readability issues, can lead too easily to returning the wrong or partially constructed values, and is generally (and correctly) frowned upon in all but the simplest of cases.

Having a universal zero value, like _ reduces the need to recognize individual entries as a zero value greatly improving the readability, but is still somewhat noisy as it must encode n zero values in the common case of return _, _, _, _, err. It is a more general proposal but, outside of returns, the use cases for a universal zero value largely only help with the case of a non-pointer struct literal. I believe the correct way to deal that is to increase the contexts in which the type of a struct literal may be elided as described in #12854

In #19642 (comment) @rogpeppe suggested the following workaround:

func f() (int, string, *T, Struct, error) {
  fail := func(err error) (int, string, *T, Struct, error) {
    return 0, "", nil, Struct{}, err
  }
  // ...
  if err != nil {
    return fail(err)
  }
  // ...
}

This has the benefit of introducing nothing new to the language. It reduces the annoyances caused by writing and editing the return values by creating a single place to write/edit the return values. It helps a lot with the reading but still has some boilerplate to read and take in. However, this pattern could be sufficient.

This proposal would complicate the grammar for the return statement and hence the go/ast package so it is not backwards compatible in the strict Go 1 sense, but as the construction is currently illegal and undefined it is compatible in the the Go 2 sense.

Possible restrictions

Only allow a single value other than ....

Only allow it on the left (return ..., err).

Do not allow it in the middle (return 1, ..., err).

While these restrictions are likely how it would be used in practice anyway, I don't see a need for the limitations. If it proves troublesome in practice it could be flagged by a linter.

Possible generalizations

Allow it to replace zero or more values making the below legal:

func f() error {
  // ...
  if err != nil {
    return ..., err
  }
  // ...
}

This would allow easier writing and editing but hurt the readability by implying that there were other possible returns. It would also make this non-sequitur legal: func noop() { return ... }. It does not seem worth it.

Allow ... in assignments. This would allow resetting of many variables to zero at once like a, b, c = ... (but not var a, b, c = ... or a, b, c := ... as their is no type to deduce) . In this case I believe the explicitness of the zero values is more a boon than an impediment. This is also far less common in actual code than a return returning multiple zero values.

[OneOfOne]

Why not named returns? func f() (i int, ss string, t *T, s Struct, err error) {}

[rogpeppe]

I like this idea, but I can't think of any place where I'd use it for anything other than filling in all but the last value. Given that, I think one could reasonably make it a little less general and allow only this form (allowing any expression instead of err, naturally)

return ...err

Note the lack of comma. I'm not sure whether it's better with a space before "err" or not.

[jimmyfrasche]

@OneOfOne Naked returns are fine for very short functions but are harder to scale: it gets too easy to accidentally return partially constructed values or the wrong err because of shadowing. Other than that, or maybe because of that, I like the explicit syntax better. A naked return says "go up to the function signature to see what can be returned, then trace through the code to see what actually gets returned here" whereas return ..., err says "I'm returning a bunch of zero values and the value bound to the err in scope"

@rogpeppe that was the original syntax proposed that I based this proposal off of. I don't like it because it appears to be a spread operator common in dynamic languages so it's a bit confusing. Having the comma makes it superficially more similar to "⋯" in mathematical writing and with like purpose. I agree that this would almost always be used as return ..., last where last is a bool or error and sometimes as a return ..., but I don't see any particular reason to artificially limit it to that. If it's not very useful it will not get used often. Is there any particular concern that a line like return 7, ... would be more confusing or error prone than a return ..., true?

[rogpeppe]

@jimmyfrasche I just don't see that it would ever be used, and given that, the comma seems like unnecessary overhead for what would be a very commonly used idiom.

How many places in existing code can you find where eliding all but the first argument (or all but several arguments) would be useful?

[ianlancetaylor]

If we permit both ..., v and v, ... how do we justify the exclusion of ..., v, ...? Except for that fact that it is impossible to implement?

[jimmyfrasche]

@rogpeppe return ...aSlice looks too much like https://developer.mozilla.org/en-US/docs/Web/JavaScript/Reference/Operators/Spread_operator That's not a strong argument against it, but I'd still rather not see it. While I would be surprised if this proposal is accepted in any form, I certainly wouldn't be mad if that variant was the one accepted, as it is the most common case. I really don't see any justification for the limitation, personally. It would be interesting to run an analysis on go-corpus to see if there are any. I don't have the time to do that for a bit, so anyone can feel free to beat me to it.

@ianlancetaylor that's a fine point. Impossible to implement is justification enough for me. Though it would be possible to implement in some cases, where v is of a type returned only once, but then subtle and distance changes could make it suddenly ambiguous. I would note that the other two existing uses of ... are also "at most once". You can't do func(a ...string, b ...int) or append(bs, s1..., s2...) even though the first is only sometimes unambiguous and the latter would be merely inefficient (or rather non-obviously inefficient).

Another option, that I'm fairly sure is a bad idea, would be to allow "keyed returns" to work in conjunction with named returns, by analogy with keyed struct literals:

func() (a int, b string, err error) {
  // ...
  if err != nil {
    return err: err
  }
  // ...
  return b: name, a: -x // I purposely flipped the order here
}

though that would interact poorly with the semantics of the naked return. If a is non-zero what happens when I write return err: err? The answer should clearly be that only err is returned, but it's still a potential source of confusion.

[mewmew]

Counter-proposal that has been suggested elsewhere in the past (#19642). Allow _ to be used as the zero value of any type in expressions. Currently _ in var _ T = foo may be thought of as foo > /dev/null (from Effective Go). Similarly, _ in var foo T = _ could be thought of as foo < /dev/null. Then _ could be used in return statements.

E.g. given the function signature func() (int, string, *T, Struct, error):

return 0, "", nil, Struct{}, nil may be written return _, _, _, _, _

return 0, "", nil, Struct{}, err may be written return return _, _, _, _, err

return 0, "", nil, Struct{X: Y}, nil may be written return _, _, _, Struct{X: Y}, _

return 1, "", nil, Struct{}, nil may be written return 1, _, _, _, _

return 1, "a", nil, Struct{}, nil may be written return 1, "a", _, _, _

return 1, "", nil, Struct{}, err may be written return 1, _, _, _, err

return 1, "a", nil, Struct{X: Y}, err may be written return 1, "a", _, Struct{X: Y}, err

[davecheney]

When your functions has so many return values that typing them becomes a chore, that's a sign that you need to redesign your function, not the language.

[mewmew]

When your functions has so many return values that typing them becomes a chore, that's a sign that you need to redesign your function, not the language.

I think the example was given simply to provide a one of each function signature that is useful as a showcase. While not likely to see so many return values in real world code, returning the zero value of a struct using _ rather than image.Rectangle{} may improve readability; at least that's the idea of the proposal.

[jimmyfrasche]

@davecheney indeed.

My argument is that the primary benefit of the succinct syntax is that it improves the readability. If it makes it easier to type that's just a bonus.

If you see return ..., err you don't need to think about what else is being returned.

You don't need to double check for things that are suspiciously close to a zero value like return Struct{''}, err or return O, err or return nil, err (when nil has been shadowed for some reason).

It's immediately obvious that the only relevant value is err. Idioms such as

if err != nil {
  return ..., err
}

can be pattern matched by your brain as a unit without having to actually inspect anything. I'm sure we all do that now with similar blocks that contain one or more zero value-like expression. It's bitten me once or twice when I was debugging and my eye glazed over something that looked too close to a zero value making it hard to spot the obvious problem (I of course do not admit publicly to being the person who shadowed nil . . .).

I'm fine with how it is, however. This is just a potential way to make it a little bit easier.

@mewmew yes this proposal is based on a comment from that proposal (see the History section). I don't particularly see the point of the generic zero value except in the case of returns. It would solve the same problem, of course.

(I would like to be able to use {} as the zero value of any struct when its type can be deduced from the rest of the statement.)

[nigeltao]

@davecheney sometimes it's not the number of return values, it's their struct-ness. Typing return image.Rectangle{}, err can be a chore. The image.Rectangle{} is the longest but also least important part of the line.

That said, this particular proposal is not the only way to sooth that chore, as the OP noted.

[ianlancetaylor]

Reopening per discussion in #33152.

[carlmjohnson]

I am in favor of allowing only the return ..., x form. If other forms are good, they could be added later.

The main benefit of this for me is that it would let me use a dumb macro to expand ife into

if err != nil {
   return ..., err
}

Yes, a sufficiently smart IDE macro could look at the function return arguments to fill those in for me, but why not just simplify it so only the important information is emphasized?

[bradfitz]

What about letting return take 1 thing regardless of how many results the func has, as long as the assignment is unambiguous to exactly 1 of the return types?

So:

    func foo() (*T, error) {
        if fail {
             return errors.New("foo")
        }
        return new(T)       
    }

[ianlancetaylor]

Does anybody see any problems with this language change?

We think we should consider just the simple case: return ..., val1, val2 where val1 and val2 become the final results of the function. Typically, of course, this would be just return ..., err. The other cases don't seem to arise enough to worry about. This would only be permitted if there are other results; it could not be used in a function that returns only error. The omitted results would be set to the zero value, even if the result parameters are named and currently set to some non-zero value.

-- for @golang/proposal-review

[jimmyfrasche]

@bradfitz that seems like it could cause too much fun when the return signature changes in a long func or one of the return types now satisfies an interface. It would also be unusable with interface{}, though that's a bit niche.

[bradfitz]

It would also be unusable with interface{}, though that's a bit niche.

That's a feature.

[andig]

I would find this unexpected or at least unlearned when s would return the zero string:

func foo() (s string, err error) {
	s = "unexpected"
	return ..., err
}

It might make more sense to say that named return values maintain their value even with .... That would make it equivalent with

return

in that case.

[vatine]

But in this case, return ..., err is not equivalent to return s, err it is rather equivalent to return "", err and I would expect that to set the named return value.

[gregwebs]

@andig would you have this same expectation if naked returns were deprecated? If this proposal were accepted, the use case for naked returns would further diminish.

[gregwebs]

There is an interesting hack in the Golang code base to achieve return ..., err

In tls.go

func X509KeyPair(certPEMBlock, keyPEMBlock []byte) (Certificate, error) {
	fail := func(err error) (Certificate, error) { return Certificate{}, err }
...
			return fail(errors.New("tls: failed to find any PEM data in certificate input"))

I wouldn't have though this possible until very recently when I learned that Go can accept multiple return values from one function as arguments to another without intermediate destructuring.

[gregwebs]

@ianlancetaylor can we get this proposal for return ..., err submitted as a Github discussion or otherwise move it through the proposal process? I am happy to help in any way I can.

[ianlancetaylor]

Moving to main proposal process.

[robpike]

I favor the direction (mentioned in various ways above; this is not an original idea) that we instead provide a unified way to express the zero of any type. And for my money, the best way to do that is with a predefined identifier called "zero". It would overlap with some uses of "nil", and we'd need to decide what to do about that (I say, nothing), but on the other hand we can open a path to a less confusing model for nil.

In other words, rather than say, "return values with zeros are hard so let's fix them, we say "zero values should be easier to use so they work well everywhere, including in return statements".

[andig]

The obvious zero values seems to be _?

[cespare]

I think that some variant of the original proposal here is better than adding zero.

For one, zero only partially mitigates the annoyances of return sites. Yes, zero is easier to type and read than image.Rectangle{}, but it's longer than "" or 0 or nil. Additionally, it doesn't solve the problem that changing the function signature leads to a diff where all the returns change (now from e.g. return zero, err to return zero, zero, err). With return ..., err, adding another return parameter wouldn't change most err != nil return sites.

Furthermore, except when writing returns, I've rarely wanted for a general-purpose way to write the zero value. In most other contexts, the zero value is close to hand without needing to name it at all. For example:

var s string // I get the zero value without mentioning it
x := struct{X int, Y float64}{Y: 3.5} // I get the zero value for X by simply eliding it

Perhaps there are cases I'm not thinking of (maybe it's more common with generics?) but I just can't think of many places I'd use zero other than returns. And so if zero mainly improves return, rather than being a broadly useful tool, it seems better to go with a more comprehensive and targeted improvement to return instead.

[atdiar]

I would tend to agree with @robpike

The issue imho is overloading a symbol such as ... or _. Now a beginner might see the same symbol in different contexts and that will make it confusing.

Also, I don't find the overall idea that compelling from my own experience writing Go. And that's actually what makes it moot: at best, it just makes things easier to write by handwaving some information.
We don't know if that might not lead to mistakes when changing function signatures.
But more importantly, it makes the code less consistent with several types of returns: fully specified and elided.
Recognizing immutable landmark idioms is easier.

On the topic of the zero identifier, that could help normalize things if writing a zero returning function is not enough (such as:

func Zero[T any] (v ...T) T{
   var z T
   return z
} 

)

[gregwebs]

Let's take a step back and clarify what problem we are solving. This proposal is trying to solve the issue of letting the programmer communicate that they are just returning an error. It can be used slightly more generally, but that is a less common use case.

The main issue is that Go forces us to write down zero values to return an error because we re-use a more general feature to return multiple return values. So the Go language is not optimized for this specific use case, which in some programs accounts for the majority of returns. That obscures readability because when reading code I need to recognize that a zero value is actually being returned at every return site.

As @cespare stated, having a generic zero value is not a problem that needs to be more solved more generally. But it also doesn't solve this specific proposal. I don't want an easier way to write down zero- I want to not have to write them at all.

However, let's look at a generic zero as a pragmatic solution of using _ as a generic zero:

• It incompletely but pragmatically fixes most of the readability of code issue (I can skip over _ when reading it)
• It incompletely but pragmatically fixes writing the code (I can write _ much quicker than auto-completing a struct name).
• It produces noisy diffs- if I make a one-line change at the end of my function to return an additional value, I now have to update all the error return sites

If there is a choice between adding _ or even zero as a generic zero, or doing nothing at all, I would gladly take the generic zero. However, if we compare a generic zero to the existing proposed solution, the existing proposed solution cleanly solves all the above three points of reading, writing, and diffing.

If we are worried we are solving this issue too specifically, that is certainly a valid concern. If we look around at other languages (for example Rust) we will see that they have solved this problem more generally with tagged unions, a powerful feature to reduce defects by properly modeling state, and the missing feature of Go for anyone that has programmed with them. If this feature were added, it would be easy for end users to define a Result type just as Rust has and solve this issue on their own.

type Result[T any] enum {
  Ok T
  Err error
}

Writing return Err(err) is an optimal solution. The problem is that for the caller of that function, they have to de-structure the Result and deal with the error. Rust has added a special error return operator ?, but it isn't clear that this is an optimal approach. De-structing a tagged union with the existing Go switch statement would be fine for most tagged unions, but the end result for the common case of returning errors will be more verbose. What Golang does now with de-structing the return immediately works pretty well for the common case, and it is not clear how to achieve the same when returning a proper Result type.

Will Go be able to add tagged unions or a Result type soon? I hope so. In the next 5 years will Go do this and solve the issue of how to conveniently deal with the result in the calling function? I am doubtful. In the mean time, it would be great if we could add ... and dramatically improve the readability and diffing of the functions we are writing every day.

[carlmjohnson]

Parts of #35966 are about adding zero as the predeclared identifier for the zero value of any type. (And all of #52307, which was closed as a dupe.) I still think it's a good idea.

[carlmjohnson]

Will Go be able to add tagged unions or a Result type soon? I hope so. In the next 5 years will Go do this and solve the issue of how to conveniently deal with the result in the calling function? I am doubtful. In the mean time, it would be great if we could add ... and dramatically improve the readability and diffing of the functions we are writing every day.

If Go programmers move towards using result types instead of multiple returns, then it will be a shame that we added special syntax just for doing a thing that modern Go code won't do anymore.

[gregwebs]

If Go programmers move towards using result types instead of multiple returns, then it will be a shame that we added special syntax just for doing a thing that modern Go code won't do anymore.

I am an optimist, I think Go will figure out how to add tagged unions eventually now that they have Generics. But to be a realist or just a pragmatist, the existing error handling works pretty well other than what is pointed out in this proposal and it is a much bigger change to make it work with a result type- I don't think it is realistic to think that Go will ever make that change.

The impact of improving Zero values

For the few times that I have needed a generic zero, and several of the uses pointed out in those tickets, defining func Zero[T any] (v T) T and similarly IsZero in the standard library somewhere would suffice, and personally I am okay with just defining it in my own code. I point this out not because I am at all against adding zero. My point is that even though improving zero values seems like a larger issue, it's actual impact on lines of code is much smaller- for me personally I would probably be returning errors 100x more than the times then where zero would improve code readability (== zero could be used in place of == nil, but I don't think that improves readability). The proposal here addresses what exists in the majority of functions that I work with. So let's not have zero values take the driver seat since trying to use them as a solution to the issue at hand is sub-optimal with comparatively noisier lines and diffs.

[zephyrtronium]

An option would be to add zero for #35966 and change this proposal to return zero..., err. Going further, we could say that (to use the spec's EBNF) 'return' {PrimaryExpr ','} PrimaryExpr '...' {',' PrimaryExpr} expands to lexically duplicating the PrimaryExpr preceding ... to all output parameters not set positionally. So, something like this:

func() (string, int, float64, complex64, error) {
    return "", 1..., nil
}()

would return string(""), int(1), float64(1), complex64(1), error(nil). I don't think that would be especially useful outside of returning zero values, but it strikes me as a more orthogonal approach.

[gregwebs]

@zephyrtronium nice suggestion:return zero..., err would satisfy the use case of "I just want to return an error" equally well.

However, I don't think we want to blur the semantics of x...- it currently expects x to be an array.

It would kind of make sense to use both dot forms at once! return (...zero)..., err where ...zero used at the value level means an array of zero values just as ...param means an array of type param at the type level. return x... would work similar to the existing Go ... to expand the array, but also set the length. But I don't think constructing arrays with ...x is that useful otherwise, and anyways that's too many dots for my taste!

Another approach would be to pre-define an array of zeros as zeros, so we would have return zeros.... But again, I don't think zeros would be very useful outside of this use case.

[gregwebs]

I opened a proposal to not require ... and instead just allow return err when there are multiple return values. I think this new proposal makes sense and avoids bike-shedding about several things in this proposal. At least this new form will seem significantly better or worse to some, so it seemed worthy or a separate proposal. I don't care what solution we come up with as long as Go will let me just return an error :)

[rsc]

This proposal has been added to the active column of the proposals project
and will now be reviewed at the weekly proposal review meetings.
— rsc for the proposal review group

[rsc]

If you have a function with things like

return 0, time.Time{}, nil, err

then the original proposal would let you write

return ..., err

which is admittedly much nicer, especially when you decide that third result wasn't needed and don't have to update all the return statements to drop the ", nil".

The "zero" suggestion doesn't help as much:

return zero, zero, zero, err

still has to get updated.

[narqo]

If you have a function with things like

return 0, time.Time{}, nil, err

I looked through the discussion above and haven't noticed if we attempted to survey the existing corpus of Go code, for if such function is that widely used, so it's worth overloading the ... and adding a new meaning to the construction.

My, very naive, guess is that functions like below are what dominate, and cause the main unsatisfaction

// 1
return SomeStruct{}, err

// 2
return SomeStruct{}, false

With that, IMHO, allowing return _, _ would solve the majority of the annoying cases.

[rsc]

Placing on hold for more thought and later discussion.

[rsc]

Placed on hold.
— rsc for the proposal review group

[mvndaai]

Any updates? I am still hoping for this any of these changes to make the return boilerplate inside an if err != nil cleaner. I have sometimes returned a *Struct{} instead of Struct{} so my returns can all be nil, err but that is a misuse.

---

To summarize the whole conversation:

The problem is with a return type that has multiple returns like (package.Struct{}, bool, error). Which is distracting to both type and read when a function has many error checks that end up looking like this.

if err != nil {
    return package.Struct{}, false, err
}

The original proposal was to allow ..., followed by any values.

return ..., err

Discussions on this on this included allowing ..., true, nil, the opposite package.Struct{}, ..., an empty .... The issue revealed is that it would make sense to allow ..., true, ... but that would be impossible for the compiler if there were multiple of the same return type in the middle like (any, any, any, error).

That led to the discussion of allowing zero value in the returns. Either as

return _, _, err

Or create a new Go builtin zero, like these other proposals #35966 and #52307. The zero would work as any zero value. Separating it from _ avoids confusion for people and the compiler.

return zero, zero, err

The complaint with a zero is if your return signature changes all the error lines also need to change.

---

I would be happy with any of them! I personally like the zero builtin because I would also use it in table unit testing.

[carlmjohnson]

I have sometimes returned a *Struct{} instead of Struct{} so my returns can all be nil, err but that is a misuse.

I was guilty of that last week as well. 😅