bytes: add Replacer

[bep]

The strings.Replacer is a real race car.

It would be very useful to have a bytes.Replacer with similar feature set.

[bradfitz]

If we do this, we should consider whether we can write it once with type T, and use go generate to spit out the strings and bytes versions. I don't know if that'll be more or less work.

[rsc]

Yes, we should add Replacer. I do not believe we should use go generate, though. Generated code is harder to work with and debug, and this will need debugging. It's not like a String method.

[josharian]

I want this too. One immediate use would be reducing allocs in the linker.

An API question: Should Replacer.Replace modify the input (a la append) or always return a copy? The former could avoid allocs in cases in which the length does not grow or the input has sufficient capacity, but it means callers have to be more careful.

[bradfitz]

I would say it should be allowed to re-use the passed-in memory. Otherwise there's little advantage over using strings.Replacer instead. That said, almost everything in bytes already is documented as returning a copy, which is a little sad.

[bep]

I vote for modifying the input.

[zombiezen]

Couldn't the strings.Replacer implementation use the bytes.Replacer implementation? There's a potential slowdown for WriteString, but it might make it simpler.

(Also, has the window for implementing this in Go 1.5 passed? I'd like to take a shot at it.)

[bradfitz]

No, I wouldn't write the strings version in terms of the bytes version. One deals with runes and one deals with bytes, in addition to the major string/[]byte representation difference.

This is still tagged Go 1.5, so perhaps if you do it soon it can still go in.

[zombiezen]

Makes sense, but after reading through the code, it appears that strings doesn't operate on runes. The representation difference is still a good argument for keeping them distinct.

[zombiezen]

This is coming along; I'm hoping to have a CL ready this week. One question on API: should the arguments of bytes.NewReplacer be strings or byte slices? strings seem easier to work with (and other functions in the bytes package use strings for character sets), but I can also see the argument for byte slices.

[bradfitz]

This is too late for Go 1.5 at this point.

[zombiezen]

Drat. I'll keep a package going at https://github.com/zombiezen/go-bytesreplacer as I work on the implementation.

[gopherbot]

CL https://golang.org/cl/13797 mentions this issue.

[rsc]

If Replacer is added it needs to mimic strings.Replacer. In particular it needs to return a copy of the input, just like bytes.Replace and bytes.Map do. Obviously one can run faster by rewriting the input slice, but this one function shouldn't be different from the design of the rest of the package.

[bradfitz]

Then it's useless and we should just close this bug.

Somebody should just maintain a useful version on Github.

[rsc]

FWIW, I disagree that it's useless.

[bradfitz]

I say it's useless because if you didn't mind copies, you could just use strings.NewReplacer and do the []byte(string) and string([]byte) copies yourself before/after strings.NewReplacer. Having bytes.NewReplacer that just does those conversions for you (but with the added huge negative of all that mostly-duplicated code) buys you very little.

[zombiezen]

For those interested, I just submitted this as go4.org/bytereplacer. Thanks @bradfitz!

[rsc]

I say it's useless because if you didn't mind copies, you could just use strings.NewReplacer and do the []byte(string) and string([]byte) copies yourself before/after strings.NewReplacer. Having bytes.NewReplacer that just does those conversions for you (but with the added huge negative of all that mostly-duplicated code) buys you very little.

By that argument we should delete the entire bytes package.

[bradfitz]

By that argument we should delete the entire bytes package.

No, by that argument, we should delete only some of the bytes package. (if only we could!)

Functions in package bytes returning bools or ints such as Compare, Contains, Count, Equal, EqualFold, HasPrefix, HasSuffix, Index, IndexAny, IndexBytes, IndexFunc, etc are fine. The mostly-useless ones are the ones returning []byte copies.

We have to keep them for compatibility, but there's no need to add more.

[rsc]

We'll just have to agree to disagree.