Add Data.ByteString.Short.unconsN

[hasufell]

Fixes #524

[Bodigrim]

Is your implementation of unconsN faster than split / unpack?

myUnconsN :: Int -> ShortByteString -> Maybe ([Word8], ShortByteString)
myUnconsN n x
  | S.length x < n = Nothing
  | otherwise = Just (S.unpack y, z)
  where
    (y, z) = S.splitAt n x

[hasufell]

Is your implementation of unconsN faster than split / unpack?

It seems so:

All
  ShortByteString
    ShortByteString unpack/uncons comparison
      unpack and look at first 5 elements: OK (1.09s)
        15.3 ms ± 1.1 ms
      uncons consecutively 5 times:        OK (0.92s)
        401  μs ±  34 μs
      unconsN 5:                           OK (0.69s)
        80.4 ns ± 5.9 ns
      myUnconsN 5:                         OK (0.93s)
        49.8 μs ± 3.3 μs

[Bodigrim]

@hasufell could you please rebase?

[sjakobi]

Is this PR still needed now that unpack performance has been improved in https://github.com/haskell/bytestring/pull/526? Frankly I don't find the API particularly appealing. For example, it seems that the [Word8] returned is never empty?!

For completeness, I think we'll also end up adding unsnocN, and variants for StrictByteString and LazyByteString – a significant amount of additional code.

[Bodigrim]

It might be that benchmarks after rebase will be more favorable to split / unpack and this PR will be redundant.

[hasufell]

All
  ShortByteString
    ShortByteString unpack/uncons comparison
      unpack and look at first 5 elements: OK (0.31s)
        412  ns ±  35 ns
      uncons consecutively 5 times:        OK (0.94s)
        406  μs ±  38 μs
      unconsN 5:                           OK (0.17s)
        77.4 ns ± 5.8 ns
      unconsNViaSplit 5:                   OK (0.15s)
        33.8 ns ± 3.0 ns

All 4 tests passed (1.61s)

[Bodigrim]

This suggests that unconsNViaSplit is the fastest option, right?

[hasufell]

I guess so