Use string slice in ascii_string and binary_string

[dkuku]

It probably requires a mention that it can reference keep a reference to the parsed binary - currently bytes combinator does the same.

Example parser:

  defmodule MyParser do
    import NimbleParsec

    defparsec :ident, ascii_string([?a..?z], min: 1), debug: true
  end

expands to:

Before (master) — 6 functions, 8 clauses:

  defp ident__0(rest, acc, stack, context, line, offset) do
    ident__1(rest, [], [acc | stack], context, line, offset)
  end

  defp ident__1(<<x0, rest::binary>>, acc, stack, context, comb__line, comb__offset) when x0 >= 97 and x0 <= 122 do
    ident__2(rest, [<<x0::integer>>] ++ acc, stack, context, comb__line, comb__offset + 1)
  end

  defp ident__1(rest, _acc, _stack, context, line, offset) do
    {:error, "expected ASCII character in the range \"a\" to \"z\"", rest, context, line, offset}
  end

  defp ident__2(<<x0, rest::binary>>, acc, stack, context, comb__line, comb__offset) when x0 >= 97 and x0 <= 122 do
    ident__4(rest, [x0] ++ acc, stack, context, comb__line, comb__offset + 1)
  end

  defp ident__2(rest, acc, stack, context, line, offset) do
    ident__3(rest, acc, stack, context, line, offset)
  end

  defp ident__4(rest, acc, stack, context, line, offset) do
    ident__2(rest, acc, stack, context, line, offset)
  end

  defp ident__3(rest, user_acc, [acc | stack], context, line, offset) do
    _ = user_acc
    ident__5(rest, [List.to_string(:lists.reverse(user_acc))] ++ acc, stack, context, line, offset)
  end

  defp ident__5(rest, acc, _stack, context, line, offset) do
    {:ok, acc, rest, context, line, offset}
  end
  

After (this branch) — 3 functions, 4 clauses:

  defp ident__0(<<x0, rest::binary>> = orig, acc, stack, context, line, offset) when x0 >= 97 and x0 <= 122 do
    ident__1(rest, orig, acc, stack, context, line, offset)
  end

  defp ident__0(rest, _acc, _stack, context, line, offset) do
    {:error, "expected ASCII character in the range \"a\" to \"z\"", rest, context, line, offset}
  end

  defp ident__1(<<x0, rest::binary>>, orig, acc, stack, context, line, offset) when x0 >= 97 and x0 <= 122 do
    ident__1(rest, orig, acc, stack, context, line, offset)
  end

  defp ident__1(rest, orig, acc, stack, context, line, offset) do
    len = byte_size(orig) - byte_size(rest)
    ident__2(rest, [binary_part(orig, 0, len) | acc], stack, context, line, offset + len)
  end

  defp ident__2(rest, acc, _stack, context, line, offset) do
    {:ok, acc, rest, context, line, offset}
  end

Benchmark result (I will include the benchmark code in a stacked pr that can be closed):

case              old(ns)  new(ns)  speedup  control   red.  red.diff
ascii len=1          60.7     53.6   1.132x   0.995x     13        -7
ascii len=4         105.4     62.0   1.699x   1.000x     17        -8
ascii len=8         122.1     67.7   1.803x   0.997x     21        -8
ascii len=32        393.6     97.5   4.038x   0.991x     48       -11
ascii len=128      1017.4    198.1   5.135x   1.018x    204       -71
ascii len=1024     6879.2   1177.2   5.844x   1.008x   1998      -969
opt len=1            67.3     44.4   1.516x   0.993x     13        -7
opt len=128        1007.2    258.2   3.902x   1.006x    224       -91
utf8 len=8          184.1     75.5   2.439x   0.970x     21        -8
utf8 len=128       1688.9    228.4   7.395x   1.007x    229       -96
bounded 2..8        114.5     64.0   1.790x   0.996x     19        -9
repeat 200x6ch    23323.4  11937.6   1.954x   1.004x   4170     -1953
repeat 200x40ch   55509.4  19811.8   2.802x   0.999x  11896     -2824
error (min fail)     36.0     34.5   1.042x   1.000x      5         0

speedup   old/new wall clock, min-of-21. Only trust it if it clears `control`.
control   the old side timed twice. This run's noise floor.
red.diff  reductions for one call, new minus old. Exact: 0 means both revisions
          perform exactly the same work, whatever the clock says.

Assisted-By: Claude Opus 5

[josevalim]

Btw, would you be so kind to send a PR that updates CI? :D