Pointers vs. values in parameters and return values

Question

In Go there are various ways to return a struct value or slice thereof. For individual ones I've seen:

type MyStruct struct {
    Val int
}

func myfunc() MyStruct {
    return MyStruct{Val: 1}
}

func myfunc() *MyStruct {
    return &MyStruct{}
}

func myfunc(s *MyStruct) {
    s.Val = 1
}

I understand the differences between these. The first returns a copy of the struct, the second a pointer to the struct value created within the function, the third expects an existing struct to be passed in and overrides the value.

I've seen all of these patterns be used in various contexts, I'm wondering what the best practices are regarding these. When would you use which? For instance, the first one could be ok for small structs (because the overhead is minimal), the second for bigger ones. And the third if you want to be extremely memory efficient, because you can easily reuse a single struct instance between calls. Are there any best practices for when to use which?

Similarly, the same question regarding slices:

func myfunc() []MyStruct {
    return []MyStruct{ MyStruct{Val: 1} }
}

func myfunc() []*MyStruct {
    return []MyStruct{ &MyStruct{Val: 1} }
}

func myfunc(s *[]MyStruct) {
    *s = []MyStruct{ MyStruct{Val: 1} }
}

func myfunc(s *[]*MyStruct) {
    *s = []MyStruct{ &MyStruct{Val: 1} }
}

Again: what are best practices here. I know slices are always pointers, so returning a pointer to a slice isn't useful. However, should I return a slice of struct values, a slice of pointers to structs, should I pass in a pointer to a slice as argument (a pattern used in the Go App Engine API)?

As you say, it really depends on the use case. All are valid depending on the situation - is this a mutable object? do we want a copy or pointer? etc. BTW you didn't mention using new(MyStruct) :) But there is no difference really between the different methods of allocating pointers and returning them. — Not_a_Golfer
That is literally over engineering. Structs must be pretty large that returning a pointer makes your program faster. Just don't bother, code, profile, fix if useful. — Volker
There is only one way to return a value or a pointer, and that is to return a value, or a pointer. How you allocate them is a separate issue. Use what works for your situation, and go write some code before you worry about it. — JimB
BTW just out of curiousity I benchamrked this. Returning structs vs. pointers seems to be roughly the same speed, but passing pointers to functions down the lines is siginificantly faster. Although not on a level it would matter — Not_a_Golfer
@Not_a_Golfer: I would assume that's just bc allocation is done outside the function. Also benchmarking values vs pointers depends on the size of the struct and memory access patterns after the fact. Copying cache-line sized things is as fast as you can get, and the speed of dereferencing pointers from CPU cache is much different than dereferencing them from main memory. — JimB

twotwotwo twotwotwo · Accepted Answer · 2014-05-08T20:34:08

tl;dr:

Methods using receiver pointers are common; the rule of thumb for receivers is, "If in doubt, use a pointer."
Slices, maps, channels, strings, function values, and interface values are implemented with pointers internally, and a pointer to them is often redundant.
Elsewhere, use pointers for big structs or structs you'll have to change, and otherwise pass values, because getting things changed by surprise via a pointer is confusing.

One case where you should often use a pointer:

Receivers are pointers more often than other arguments. It's not unusual for methods to modify the thing they're called on, or for named types to be large structs, so the guidance is to default to pointers except in rare cases.
- Jeff Hodges' copyfighter tool automatically searches for non-tiny receivers passed by value.

Some situations where you don't need pointers:

Code review guidelines suggest passing small structs like type Point struct { latitude, longitude float64 }, and maybe even things a bit bigger, as values, unless the function you're calling needs to be able to modify them in place.
- Value semantics avoid aliasing situations where an assignment over here changes a value over there by surprise.
- It's not Go-y to sacrifice clean semantics for a little speed, and sometimes passing small structs by value is actually more efficient, because it avoids cache misses or heap allocations.
- So, Go Wiki's code review comments page suggests passing by value when structs are small and likely to stay that way.
- If the "large" cutoff seems vague, it is; arguably many structs are in a range where either a pointer or a value is OK. As a lower bound, the code review comments suggest slices (three machine words) are reasonable to use as value receivers. As something nearer an upper bound, bytes.Replace takes 10 words' worth of args (three slices and an int). You can find situations where copying even large structs turns out a performance win, but the rule of thumb is not to.
For slices, you don't need to pass a pointer to change elements of the array. io.Reader.Read(p []byte) changes the bytes of p, for instance. It's arguably a special case of "treat little structs like values," since internally you're passing around a little structure called a slice header (see Russ Cox (rsc)'s explanation). Similarly, you don't need a pointer to modify a map or communicate on a channel.
For slices you'll reslice (change the start/length/capacity of), built-in functions like append accept a slice value and return a new one. I'd imitate that; it avoids aliasing, returning a new slice helps call attention to the fact that a new array might be allocated, and it's familiar to callers.
- It's not always practical follow that pattern. Some tools like database interfaces or serializers need to append to a slice whose type isn't known at compile time. They sometimes accept a pointer to a slice in an interface{} parameter.
Maps, channels, strings, and function and interface values, like slices, are internally references or structures that contain references already, so if you're just trying to avoid getting the underlying data copied, you don't need to pass pointers to them. (rsc wrote a separate post on how interface values are stored).
- You still may need to pass pointers in the rarer case that you want to modify the caller's struct: flag.StringVar takes a *string for that reason, for example.

Where you use pointers:

Consider whether your function should be a method on whichever struct you need a pointer to. People expect a lot of methods on x to modify x, so making the modified struct the receiver may help to minimize surprise. There are guidelines on when receivers should be pointers.
Functions that have effects on their non-receiver params should make that clear in the godoc, or better yet, the godoc and the name (like reader.WriteTo(writer)).
You mention accepting a pointer to avoid allocations by allowing reuse; changing APIs for the sake of memory reuse is an optimization I'd delay until it's clear the allocations have a nontrivial cost, and then I'd look for a way that doesn't force the trickier API on all users:
1. For avoiding allocations, Go's escape analysis is your friend. You can sometimes help it avoid heap allocations by making types that can be initialized with a trivial constructor, a plain literal, or a useful zero value like bytes.Buffer.
2. Consider a Reset() method to put an object back in a blank state, like some stdlib types offer. Users who don't care or can't save an allocation don't have to call it.
3. Consider writing modify-in-place methods and create-from-scratch functions as matching pairs, for convenience: existingUser.LoadFromJSON(json []byte) error could be wrapped by NewUserFromJSON(json []byte) (*User, error). Again, it pushes the choice between laziness and pinching allocations to the individual caller.
4. Callers seeking to recycle memory can let sync.Pool handle some details. If a particular allocation creates a lot of memory pressure, you're confident you know when the alloc is no longer used, and you don't have a better optimization available, sync.Pool can help. (CloudFlare published a useful (pre-sync.Pool) blog post about recycling.)

Finally, on whether your slices should be of pointers: slices of values can be useful, and save you allocations and cache misses. There can be blockers:

The API to create your items might force pointers on you, e.g. you have to call NewFoo() *Foo rather than let Go initialize with the zero value.
The desired lifetimes of the items might not all be the same. The whole slice is freed at once; if 99% of the items are no longer useful but you have pointers to the other 1%, all of the array remains allocated.
Moving values around might cause you performance or correctness problems, making pointers more attractice. Notably, append copies items when it grows the underlying array. Pointers you got before the append point to the wrong place after, copying can be slower for huge structs, and for e.g. sync.Mutex copying isn't allowed. Insert/delete in the middle and sorting similarly move items around.

Broadly, value slices can make sense if either you get all of your items in place up front and don't move them (e.g., no more appends after initial setup), or if you do keep moving them around but you're sure that's OK (no/careful use of pointers to items, items are small enough to copy efficiently, etc.). Sometimes you have to think about or measure the specifics of your situation, but that's a rough guide.

Pointers vs. values in parameters and return values

5 Answers