bytes, strings: add iterator forms of existing functions

[rsc]

We propose to add the following functions to package bytes and package strings, to allow callers to iterate over these results without having to allocate the entire result slice. This text shows only the string package form.

This is one of a collection of proposals updating the standard library for the new 'range over function' feature (#61405). It would only be accepted if that proposal is accepted. See #61897 for a list of related proposals.

---

Iterating over lines is an incredibly common operation that we’ve resisted adding only because we didn’t want to encourage allocation of a potentially large slice. Iterators provide a way to finally add it.

// Lines returns an iterator over the newline-terminated lines in the string s.
// The lines yielded by the iterator include their terminating newlines.
// If s is empty, the iterator yields no lines at all.
// If s does not end in a newline, the final yielded line will not end in a newline.
func Lines(s string) iter.Seq[string] {
	return func(yield func(string)bool) bool {
		for s != "" {
			var line string
			if i := strings.Index(s, "\n"); i >= 0 {
				line, s = s[:i+1], s[i+1:]
			} else {
				line, s = s, ""
			}
			if !yield(line) {
				return false
			}
		}
		return true
	}
}

Iterating over bytes in a string is common and too difficult, since range ranges over runes. This function will inline to the obvious for loop (because we will make sure it does):

// Bytes returns an iterator over bytes in s, yielding both the index and the byte.
func Bytes(s string) iter.Seq2[int, byte] {
	return func(yield func(int, byte) bool) bool {
		for i := range len(s) {
			if !yield(i, s[i]) {
				return false
			}
		}
		return true
	}
}

Iterating over runes is served by a regular range loop, but like slices.All and maps.All, it could be useful as an input to other iterator adapters. The name is Runes, not Seq or All, so that its clear at call sites what is being iterated over (runes not bytes).

// Runes returns an iterator over bytes in s, yielding both the start index and the rune.
func Runes(s string) iter.Seq2[int, rune] {
	return func(yield func(int, rune) bool) bool {
		for i, c := range s {
			if !yield(i, c) {
				return false
			}
		}
		return true
	}
}

Similar to Lines, there should be iterator forms of Split, Fields, and Runes, to avoid requiring the allocation of a slice when the caller only wants to iterate over the individual results. If we were writing the library from scratch, we might use the names Split, Fields, and Runes for the iterator-returning versions, and code that wanted the full slice could use slices.Collect. But that's not an option here, so we add a distinguishing Seq suffix. We do not expect that new functions will use the Seq suffix. For example the function above is Lines, not LinesSeq.

// SplitSeq returns an iterator over all substrings of s separated by sep.
// The iterator yields the same strings that would be returned by Split(s, sep),
// but without constructing the slice.
func SplitSeq(s, sep string) iter.Seq[string] {
	if sep == "" {
		return runeSplitSeq(s)
	}
	return func(yield func(string)bool) bool {
		for {
			i := strings.Index(s, sep)
			if i < 0 {
				break
			}
			frag := s[:i]
			if !yield(frag) {
				return false
			}
			s = s[i+len(sep):]
		}
		return yield(s)
	}
}

func runeSplitSeq(s string) iter.Seq[string] {
	return func(yield func(string)bool) bool {
		for s != "" {
			_, size := utf8.DecodeRuneInString(s)
			if !yield(s[:size]) {
				return false
			}
			s = s[size:]
		}
	}
}

// SplitAfterSeq returns an iterator over substrings of s split after each instance of sep.
func SplitAfterSeq(s, sep string) iter.Seq[string]

// FieldsSeq returns an iterator over substrings of s split around runs of
// whitespace characters, as defined by unicode.IsSpace. ...
func FieldsSeq(s string) iter.Seq[string]

// FieldsFuncSeq returns an iterator over substrings of s split around runs of
// Unicode code points satisfying f(c). ...
func FieldsFuncSeq(s string, f func(rune) bool) iter.Seq[string]

[bcmills]

Iterating over runes is served by a regular range loop

Will Bytes and Runes be added to the bytes package as well? Runes in particular is not entirely trivial today for byte-slices.

[jimmyfrasche]

Should Lines also yield the 1-based line number? It's not always needed but it'd be easy enough to ignore when you don't.

[earthboundkid]

Iterating over bytes in a string is common and too difficult, since range ranges over runes.

Is i, c := range []byte(s) too difficult? I would say the advantage is that strings.Bytes is more discoverable, but the []byte() version is preferred when you don't need an iterator for some reason.

[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

[DeedleFake]

Iterating over bytes in a string is common and too difficult, since range ranges over runes.

Is i, c := range []byte(s) too difficult? I would say the advantage is that strings.Bytes is more discoverable, but the []byte() version is preferred when you don't need an iterator for some reason.

That relies on specialized optimizations to avoid an allocation and a copy. I'm not sure if those optimizations exist right now. The current best way to do it, as far as I know, is

for i := 0; i < len(s); i++ {
  c := s[i]
  // ...
}

I don't think that's particularly difficult, personally, but I think that strings.Bytes() could still be useful as a feed in to a transformative iterator.

[ianlancetaylor]

Similar to Lines, there should be iterator forms of Split, Fields, and Runes

Unless I misunderstand I think you just mean Split and Fields. And there is another misplaced reference to Runes a few lines down.

[benhoyt]

I like this overall. Two things I wonder about:

1. It seems to me that Lines should trim the line ending from the string. That's how the existing similar tool bufio.ScanLines works, and it's almost always what you want, otherwise you have to call strings.TrimSpace or strings.TrimRight before you use the string. It seems to me for the few cases you really need the line ending (or the ability to rebuild the exact input text) you can use strings.Index manually.
2. I wonder if SplitSeq and FieldsSeq will pay their way. I often need the length of the split result, for example with ad-hoc parsing of separator-separated values, you often want to know the number of columns/fields. (Though I admit that since strings.Cut, that takes care of quite a few previous use cases.) Should we do some analysis of this in real-world code to determine whether it's worth adding these?

[andig]

Is i, c := range []byte(s) too difficult?

Don‘t think so. And shouldn‘t there be only one way of doing things in go?

[jimmyfrasche]

You can pass iterators to other functions instead of just immediately using it in a range so being able to make iterators for built in things is still useful.

[aarzilli]

ISTM that strings.Lines encourages people to write:

buf, _ := os.ReadFile(
for _, line := range strings.Lines(string(buf)) { ... }

I'd rather see an iterator version of bufio.Scanner.

[Merovius]

@aarzilli I think it's pretty trivial to add an All() method to bufio.Scanner. strings.Lines would still be useful. Note that s := bufio.NewScanner(strings.NewReader(str)); for _, line := s.All() { … } requires to copy data around a bunch (because of how io.Reader works), while for _, line := strings.Lines(str) { … } doesn't copy anything.

[willfaught]

Bytes(s string) iter.Seq2[int, byte]

I wonder if it's necessary to return the index for bytes, since it's trivial to count them yourself. I guess it's nice to mirror the for/range statement.

func SplitSeq(s, sep string) iter.Seq[string]

Should this return the start index of the string too?

Similar to Lines, there should be iterator forms of Split, Fields, and Runes

Should we have SplitAfterNSeq and SplitNSeq too? If not, why? These are the only ones left out.

I wonder if SplitSeq and FieldsSeq will pay their way. I often need the length of the split result, for example with ad-hoc parsing of separator-separated values, you often want to know the number of columns/fields. (Though I admit that since strings.Cut, that takes care of quite a few previous use cases.) Should we do some analysis of this in real-world code to determine whether it's worth adding these?

If you need to know the length, the existing functions can handle that case.

[earthboundkid]

Join would be useful to have. For example, just now I wanted to make a String() method for a named slice. The easy way to do it would be strings.JoinAll(iterutil.Map(slices.Values(s), Element.String), ", ").

[rsc]

Finishing this proposal discussion is blocked on #61405.

[gopherbot]

Change https://go.dev/cl/558735 mentions this issue: bytes, strings: add SplitSeq, SplitAfterSeq, FieldsSeq, FieldsFuncSeq

[rsc]

Have all remaining concerns about this proposal been addressed?

The proposal is to add these to both bytes and strings:

// Lines returns an iterator over the newline-terminated lines in the string s.
// The lines yielded by the iterator include their terminating newlines.
// If s is empty, the iterator yields no lines at all.
// If s does not end in a newline, the final yielded line will not end in a newline.
func Lines(s string) iter.Seq[string]

// SplitSeq returns an iterator over all substrings of s separated by sep.
// The iterator yields the same strings that would be returned by Split(s, sep),
// but without constructing the slice.
func SplitSeq(s, sep string) iter.Seq[string]

// SplitAfterSeq returns an iterator over substrings of s split after each instance of sep.
// The iterator yields the same strings that would be returned by SplitAfter(s, sep),
// but without constructing the slice.
func SplitAfterSeq(s, sep string) iter.Seq[string]

// FieldsSeq returns an iterator over substrings of s split around runs of
// whitespace characters, as defined by unicode.IsSpace.
// The iterator yields the same strings that would be returned by Fields(s),
// but without constructing the slice.
func FieldsSeq(s string) iter.Seq[string]

// FieldsFuncSeq returns an iterator over substrings of s split around runs of
// Unicode code points satisfying f(c).
// The iterator yields the same strings that would be returned by FieldsFunc(s),
// but without constructing the slice.
func FieldsFuncSeq(s string, f func(rune) bool) iter.Seq[string]

[fzipp]

// Lines returns an iterator over the newline-terminated lines in the string s.
// The lines yielded by the iterator include their terminating newlines.

I would not expect them to include the terminating newlines if I didn't read the documentation.

[rsc]

I would not expect them to include the terminating newlines if I didn't read the documentation.

Good reason to read the documentation! 😄

If the newlines are not included, then the user of the iterator cannot distinguish Lines("abc") from Lines("abc\n"). It is often important to know whether a file has a non-terminated final line.

[DeedleFake]

If the 'last' line does end in a newline, does Lines() yield an empty line without a newline after it? In other words, would Lines("example\n") yield just "example\n" or would it yield "example\n" and then ""?

Edit: And if it does yield that last line, how is Lines(str) any different from SplitAfterSeq(str, "\n") besides having a cleaner, shorter name?

[jimmyfrasche]

I know there are cases where it matters if the file ends with a newline or not but I don't think I have ever had to deal with that in any code I have ever written.

[earthboundkid]

I think the reason to keep the \n is then you don't have to document what happens to \r\n.

[rsc]

If the 'last' line does end in a newline, does Lines() yield an empty line without a newline after it? In other words, would Lines("example\n") yield just "example\n" or would it yield "example\n" and then ""?

It does not yield a final "".

[jub0bs]

I'm late to the party, but I noticed that the doc comments (at tip) of Lines, SplitSeq, and SplitAtfterSeq (in both the bytes and strings packages) all contain the following statement:

It returns a single-use iterator.

By inspecting the implementation of those functions, I understand why the iterators they return are single-use. For example, the push iterator returned by strings.Lines is stateful because it closes over that function's string parameter and mutates it:

func Lines(s string) iter.Seq[string] {
	return func(yield func(string) bool) {
		for len(s) > 0 {
			var line string
			if i := IndexByte(s, '\n'); i >= 0 {
				line, s = s[:i+1], s[i+1:] // <-------------
			} else {
				line, s = s, ""
			}
			if !yield(line) {
				return
			}
		}
		return
	}
}

However, I don't understand why the resulting iterators have to be single-use. I may have missed something, but I don't think their single-use nature was discussed in this proposal. What's the rationale for it?

For convenience, shouldn't iterators be reusable whenever possible/inexpensive? If reusability is indeed desired, achieving it in this case would be straightforward; for instance, for strings.Lines, it's a one-line fix:

func Lines(s string) iter.Seq[string] {
	return func(yield func(string) bool) {
		s := s // <--------- local copy
		for len(s) > 0 {
			var line string
			if i := IndexByte(s, '\n'); i >= 0 {
				line, s = s[:i+1], s[i+1:]
			} else {
				line, s = s, ""
			}
			if !yield(line) {
				return
			}
		}
		return
	}
}

(playground)

[ianlancetaylor]

We wouldn't want to do that for bytes.Lines, because the byte slice might change. And there is something to be said for consistency between strings.Lines and bytes.Lines. So, maybe we should change it, but I'm not sure.

[jub0bs]

@ianlancetaylor Good point. I guess consistency between strings.Lines and bytes.Lines should prevail over the reusability of strings iterators.

[jub0bs]

I know the proposal has already been accepted (which rejoices me), but here is some anecdotal argument in favour of strings.SplitSeq's addition: rs/cors#170

[earthboundkid]

Today I wished I had strings.JoinSeq (and could have sworn it was already in Go 1.23), so I opened #70034.

[gopherbot]

Change https://go.dev/cl/637176 mentions this issue: bytes, strings: add cross-references in docstrings

[gopherbot]

Change https://go.dev/cl/637358 mentions this issue: bytes, strings: examples for funcs added in Go 1.24

[gopherbot]

Change https://go.dev/cl/647875 mentions this issue: bytes: use "subslice" instead of "substring" in doc comments

[gopherbot]

Change https://go.dev/cl/648095 mentions this issue: [release-branch.go1.24] bytes: use "subslice" instead of "substring" in doc comments