remove intermediate list in float parse

[dkuku]

This change validates the input to Float.parse on the original binary and passes a slice of the original binary to :erlang.binary_to_float() or :erlang.binary_to_interger()

Benchmarked it for different inputs and it always uses less memory and is almost always faster. The only slower edge case is a float without a dot 1e3 where it requires an additional .0 to be added to be correctly parsed - first 2 examples in the benchmark - I think this is an edge case - the correct version should be 1.0e3.
Integer input is 2x faster - it just calculates integer * 1.0.

Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com

[dkuku]

benchmark;

Mix.install([{:benchee, "~> 1.3"}])

defmodule FloatOld do
  @moduledoc "Previous implementation: builds a reversed char list."

  def parse("-" <> binary) do
    case parse_unsigned(binary) do
      :error -> :error
      {number, remainder} -> {-number, remainder}
    end
  end

  def parse("+" <> binary), do: parse_unsigned(binary)
  def parse(binary), do: parse_unsigned(binary)

  defp parse_unsigned(<<digit, rest::binary>>) when digit in ?0..?9,
    do: parse_unsigned(rest, false, false, [digit])

  defp parse_unsigned(binary) when is_binary(binary), do: :error

  defp parse_unsigned(<<digit, rest::binary>>, dot?, e?, acc) when digit in ?0..?9,
    do: parse_unsigned(rest, dot?, e?, [digit | acc])

  defp parse_unsigned(<<?., digit, rest::binary>>, false, false, acc) when digit in ?0..?9,
    do: parse_unsigned(rest, true, false, [digit, ?. | acc])

  defp parse_unsigned(<<exp_marker, digit, rest::binary>>, dot?, false, acc)
       when exp_marker in ~c"eE" and digit in ?0..?9,
       do: parse_unsigned(rest, true, true, [digit, ?e | add_dot(acc, dot?)])

  defp parse_unsigned(<<exp_marker, sign, digit, rest::binary>>, dot?, false, acc)
       when exp_marker in ~c"eE" and sign in ~c"-+" and digit in ?0..?9,
       do: parse_unsigned(rest, true, true, [digit, sign, ?e | add_dot(acc, dot?)])

  defp parse_unsigned(rest, dot?, _, acc) do
    acc
    |> add_dot(dot?)
    |> :lists.reverse()
    |> :erlang.list_to_float()
  rescue
    ArgumentError -> :error
  else
    float -> {float, rest}
  end

  defp add_dot(acc, true), do: acc
  defp add_dot(acc, false), do: [?0, ?. | acc]
end

inputs = %{
  "int-exponent \"34e5\"" => "34e5",
  "int-exponent signed \"34e-5\"" => "34e-5",
  "int-exponent trailing \"34e5xyz\"" => "34e5xyz",
  "integer \"34\"" => "34",
  "integer \"1234567\"" => "1234567",
  "decimal \"34.25\"" => "34.25",
  "decimal \"3.14159265358979\"" => "3.14159265358979",
  "scientific \"1.5e10\"" => "1.5e10",
  "scientific \"6.022e23\"" => "6.022e23",
  "negative \"-273.15\"" => "-273.15",
  "trailing \"56.5xyz\"" => "56.5xyz",
  "long decimal" => String.duplicate("1", 40) <> "." <> String.duplicate("2", 40)
}

# Both implementations must agree before benchmarking.
for {label, input} <- inputs do
  results = [Float.parse(input), FloatOld.parse(input)]

  unless match?([r, r], results) do
    raise "mismatch for #{label}: Float=#{inspect(Enum.at(results, 0))} " <>
            "FloatOld=#{inspect(Enum.at(results, 1))}"
  end
end

IO.puts("Both implementations agree. Running Benchee...\n")

Benchee.run(
  %{
    "Float (current)" => fn input -> Float.parse(input) end,
    "FloatOld" => fn input -> FloatOld.parse(input) end
  },
  inputs: inputs,
  warmup: 1,
  time: 1,
  memory_time: 1
)

[dkuku]

results:

##### With input int-exponent "34e5" #####
Name                      ips        average  deviation         median         99th %
FloatOld               6.68 M      149.60 ns  ±4943.32%         120 ns         221 ns
Float (current)        3.61 M      277.35 ns  ±1651.06%         210 ns        1413 ns

Comparison:
FloatOld               6.68 M
Float (current)        3.61 M - 1.85x slower +127.76 ns

Memory usage statistics:

Name               Memory usage
FloatOld                  288 B
Float (current)           184 B - 0.64x memory usage -104 B


##### With input int-exponent signed "34e-5" #####
Name                      ips        average  deviation         median         99th %
FloatOld               5.72 M      174.86 ns  ±4612.21%         140 ns         201 ns
Float (current)        3.51 M      284.70 ns  ±1567.16%         231 ns        1383 ns

Comparison:
FloatOld               5.72 M
Float (current)        3.51 M - 1.63x slower +109.84 ns

Memory usage statistics:

Name               Memory usage
FloatOld                  320 B
Float (current)           184 B - 0.57x memory usage -136 B


##### With input int-exponent trailing "34e5xyz" #####
Name                      ips        average  deviation         median         99th %
FloatOld               6.53 M      153.04 ns  ±4098.75%         120 ns         200 ns
Float (current)        3.56 M      281.14 ns  ±1911.86%         210 ns        1363 ns

Comparison:
FloatOld               6.53 M
Float (current)        3.56 M - 1.84x slower +128.10 ns

Memory usage statistics:

Name               Memory usage
FloatOld                  296 B
Float (current)           192 B - 0.65x memory usage -104 B


##### With input decimal "3.14159265358979" #####
Name                      ips        average  deviation         median         99th %
Float (current)        5.56 M      179.88 ns  ±3804.16%         160 ns         300 ns
FloatOld               4.33 M      230.79 ns  ±2516.39%         191 ns         361 ns

Comparison:
Float (current)        5.56 M
FloatOld               4.33 M - 1.28x slower +50.91 ns

Memory usage statistics:

Name               Memory usage
Float (current)           128 B
FloatOld                  608 B - 4.75x memory usage +480 B


##### With input decimal "34.25" #####
Name                      ips        average  deviation         median         99th %
Float (current)        6.78 M      147.58 ns  ±5074.43%         121 ns         190 ns
FloatOld               6.25 M      159.98 ns  ±4808.09%         130 ns         251 ns

Comparison:
Float (current)        6.78 M
FloatOld               6.25 M - 1.08x slower +12.40 ns

Memory usage statistics:

Name               Memory usage
Float (current)           120 B
FloatOld                  256 B - 2.13x memory usage +136 B


##### With input integer "1234567" #####
Name                      ips        average  deviation         median         99th %
Float (current)       12.15 M       82.33 ns  ±5891.43%          60 ns         110 ns
FloatOld               6.06 M      165.04 ns  ±3652.72%         130 ns         290 ns

Comparison:
Float (current)       12.15 M
FloatOld               6.06 M - 2.00x slower +82.71 ns

Memory usage statistics:

Name               Memory usage
Float (current)           120 B
FloatOld                  384 B - 3.20x memory usage +264 B


##### With input integer "34" #####
Name                      ips        average  deviation         median         99th %
Float (current)       14.48 M       69.07 ns  ±7223.35%          50 ns          80 ns
FloatOld               7.39 M      135.26 ns  ±4570.69%         100 ns         191 ns

Comparison:
Float (current)       14.48 M
FloatOld               7.39 M - 1.96x slower +66.18 ns

Memory usage statistics:

Name               Memory usage
Float (current)           120 B
FloatOld                  224 B - 1.87x memory usage +104 B


##### With input long decimal #####
Name                      ips        average  deviation         median         99th %
Float (current)        3.01 M      331.83 ns  ±1637.08%         310 ns         471 ns
FloatOld               1.66 M      602.59 ns  ±1185.38%         541 ns         822 ns

Comparison:
Float (current)        3.01 M
FloatOld               1.66 M - 1.82x slower +270.76 ns

Memory usage statistics:

Name               Memory usage
Float (current)        0.133 KB
FloatOld                1.58 KB - 11.88x memory usage +1.45 KB


##### With input negative "-273.15" #####
Name                      ips        average  deviation         median         99th %
FloatOld               6.16 M      162.45 ns   ±523.80%         150 ns         281 ns
Float (current)        5.75 M      174.02 ns  ±3475.28%         140 ns         261 ns

Comparison:
FloatOld               6.16 M
Float (current)        5.75 M - 1.07x slower +11.57 ns

Memory usage statistics:

Name               Memory usage
FloatOld                  312 B
Float (current)           144 B - 0.46x memory usage -168 B


##### With input scientific "1.5e10" #####
Name                      ips        average  deviation         median         99th %
Float (current)        6.66 M      150.22 ns  ±3420.57%         130 ns         211 ns
FloatOld               6.55 M      152.70 ns  ±5004.33%         120 ns         181 ns

Comparison:
Float (current)        6.66 M
FloatOld               6.55 M - 1.02x slower +2.47 ns

Memory usage statistics:

Name               Memory usage
Float (current)           120 B
FloatOld                  288 B - 2.40x memory usage +168 B


##### With input scientific "6.022e23" #####
Name                      ips        average  deviation         median         99th %
Float (current)        6.32 M      158.30 ns  ±3542.89%         131 ns         220 ns
FloatOld               5.74 M      174.13 ns  ±2492.93%         150 ns         270 ns

Comparison:
Float (current)        6.32 M
FloatOld               5.74 M - 1.10x slower +15.83 ns

Memory usage statistics:

Name               Memory usage
Float (current)           120 B
FloatOld                  352 B - 2.93x memory usage +232 B


##### With input trailing "56.5xyz" #####
Name                      ips        average  deviation         median         99th %
FloatOld               6.41 M      156.02 ns  ±3831.06%         121 ns         201 ns
Float (current)        6.28 M      159.25 ns  ±1860.85%         150 ns         280 ns

Comparison:
FloatOld               6.41 M
Float (current)        6.28 M - 1.02x slower +3.23 ns

Memory usage statistics:

Name               Memory usage
FloatOld                  232 B
Float (current)           128 B - 0.55x memory usage -104 B

[dkuku]

with iodata_to_binary it is faster on the integer with exponent path:

##### With input int-exponent "34e5" #####
Name                      ips        average  deviation         median         99th %
FloatOld               6.53 M      153.04 ns  ±5814.86%         120 ns         181 ns
Float (current)        4.41 M      226.58 ns  ±3370.24%         181 ns         340 ns

Comparison:
FloatOld               6.53 M
Float (current)        4.41 M - 1.48x slower +73.53 ns

Memory usage statistics:

Name               Memory usage
FloatOld                  288 B
Float (current)           192 B - 0.67x memory usage -96 B

**All measurements for memory usage were the same**

##### With input int-exponent signed "34e-5" #####
Name                      ips        average  deviation         median         99th %
FloatOld               5.74 M      174.30 ns  ±4524.43%         140 ns         250 ns
Float (current)        3.96 M      252.33 ns  ±2753.32%         210 ns         361 ns

Comparison:
FloatOld               5.74 M
Float (current)        3.96 M - 1.45x slower +78.03 ns

Memory usage statistics:

Name               Memory usage
FloatOld                  320 B
Float (current)           192 B - 0.60x memory usage -128 B

**All measurements for memory usage were the same**

##### With input int-exponent trailing "34e5xyz" #####
Name                      ips        average  deviation         median         99th %
FloatOld               6.48 M      154.30 ns  ±4189.96%         121 ns         201 ns
Float (current)        4.46 M      224.11 ns  ±2546.46%         190 ns         341 ns

Comparison:
FloatOld               6.48 M
Float (current)        4.46 M - 1.45x slower +69.82 ns

Memory usage statistics:

Name               Memory usage
FloatOld                  296 B
Float (current)           200 B - 0.68x memory usage -96 B

[josevalim]

💚 💙 💜 💛 ❤️