String split zero copy

[malthe]

Rendered

Implementation is here: malthe/ponyc@issue-1309...malthe:string-split-zero-copy

[jemc]

Can you make explicit the method signature of the current String.split method as it compares to the new one?

[SeanTAllen]

I need convincing. I can see having another split method that does this. We have/had that at Sendence. However changing split to only work on vals is not a good thing. Its perfectly reasonable for me to want to be able to split a ref. This change would disallow that. Rather than changing the existing method, I believe what you want is a new method.

[malthe]

I suppose with subtyping it would be quite easy to provide two implementations of String.split – one for val and one for ref.

[SeanTAllen]

What subtyping are you referring to @malthe?

[malthe]

I was thinking that with #62 it might be possible to "overload" String.split with two implementations.

If that's something we want?

An alternative is of course to go with different names but that seems a bit wrong to me – like Hungarian notation almost.

I think ergonomics should be a concern here as well as obviously performance.

[jemc]

I wonder how the performance of (String box).clone -> (String val).split compares to the current case of (String box).split. It's possible that the clone.split would actually be more performant, since it involves one large buffer allocation instead of many small buffer allocations. If that ends up being true, I'd think that there is not much justification to keep (String box).split around.

It would be interesting to see some benchmarks.

[SeanTAllen]

@jemc that would be an interesting benchmark. did you mean:

(String box).split in the following?

I'd think that there is not much justification to keep (String box).clone around.

[jemc]

@SeanTAllen yes, thanks for the correction - I've corrected the comment.

[malthe]

The benchmark results are:

split               20    82373935 ns/op
split-zc            30    44659338 ns/op

For what it's worth, ponylang/ponyc#1335 actually runs about 1% faster than "split".

Program code:

use "ponybench"

actor Main
  let length: USize = 1024 * 1024

  new create(env: Env) =>
    let s = recover ref String.create(length) end

    while s.size() < length do
      s.append("abcde ")
    end

    let s': String val = s.clone()
    let bench = PonyBench(env)

    bench[None]("split", recover val lambda()(s') => s'.split() end end)
    bench[None]("split-zc", recover val lambda()(s') => s'.clone().split_zc() end end)

[jemc]

@malthe thanks for doing this benchmark.

I cloned your fork and tried a few variations on the benchmark, and couldn't come up with a case where the split() wasn't slower than clone().split_zc().

With that in mind, I think there is no reason to keep a fun split because fun val split is more optimal, even when you have to clone first to get a val.

[SeanTAllen]

Closing as stale.