Benchmarking dot in map.field

The benchmark

Mix.install [:benchee]

defmodule X do

  def dot(map) do
    [
      map.x + map.y,
      map.y + map.x,
      map.x + map.y,
      map.y + map.x,
      map.x + map.y,
      map.y + map.x,
    ]
  end

  def annotated(%{} = map) do
    [
      map.x + map.y,
      map.y + map.x,
      map.x + map.y,
      map.y + map.x,
      map.x + map.y,
      map.y + map.x,
    ]
  end

  def match(%{x: x, y: y}) do
    [
      x + y,
      y + x,
      x + y,
      y + x,
      x + y,
      y + x,
    ]
  end

  defmacrop dot(map, key) do
    quote do
      :erlang.map_get(unquote(key), unquote(map))
    end
  end

  def safe_dot(map) do
    [
      dot(map, :x) + dot(map, :y),
      dot(map, :y) + dot(map, :x),
      dot(map, :x) + dot(map, :y),
      dot(map, :y) + dot(map, :x),
      dot(map, :x) + dot(map, :y),
      dot(map, :y) + dot(map, :x),
    ]
  end

  defmacrop match(map, field) do
    quote do
      %{unquote(field) => value} = unquote(map)
      value
    end
  end

  def match_every_time(map) do
    [
      match(map, :x) + match(map, :y),
      match(map, :y) + match(map, :x),
      match(map, :x) + match(map, :y),
      match(map, :y) + match(map, :x),
      match(map, :x) + match(map, :y),
      match(map, :y) + match(map, :x),
    ]
  end

end

Benchee.run(%{
  "dot" => fn -> X.dot(%{x: 1, y: 1}) end,
  "annotated" => fn -> X.annotated(%{x: 1, y: 1}) end,
  "erlang.map_get" => fn -> X.safe_dot(%{x: 1, y: 1}) end,
  "match" => fn -> X.match(%{x: 1, y: 1}) end,
  "match_every_time" => fn -> X.match_every_time(%{x: 1, y: 1}) end,
}, warmup: 1, time: 2)

And the result is:

Name                       ips        average  deviation         median         99th %
match                   7.46 M      133.99 ns ±22658.59%          90 ns         184 ns
match_every_time        7.33 M      136.44 ns ±23751.72%          93 ns         181 ns
annotated               7.22 M      138.59 ns ±23522.99%          92 ns         196 ns
erlang.map_get          6.67 M      149.82 ns ±21303.83%         105 ns         212 ns
dot                     4.24 M      235.62 ns ±16830.96%         137 ns         266 ns

Comparison:
match                   7.46 M
match_every_time        7.33 M - 1.02x slower +2.45 ns
annotated               7.22 M - 1.03x slower +4.60 ns
erlang.map_get          6.67 M - 1.12x slower +15.83 ns
dot                     4.24 M - 1.76x slower +101.63 ns

Explanation

First of all, construction like map.field get’s compiled into this by elixir compiler. This is a feature of Elixir runtime and it is caused by feature of calling functions without (). However, code map.field() gets compiled to the same expression.

case map do
  module when is_atom(module) -> apply(module, :field, [])
  %{field: value} -> value
  _ -> raise BadMapError
end

• match generates the fastest core_erlang and beam for this problem. Rule of thumb: pattern matching is always the fastest way to access the data in collections in both Erlang and Elixir.
• annotated generates almost the same core_erlang as dot version, but compiler sees that the map variable in the function body will always be a map (otherwise it wouldn’t pass the matching in args) and it optimizes the case generated by dot to just a pattern matching. So it generates the same BEAM as match version
• erlang.map_get is a very unpopular way to access the key of a map, and it is the third way to do so, added in one of the latest OTP version. It is actually a separate BIF and it is slower because it performs a map type check on every call
• dot is the slowest one, because it performs a type check and a jump on every key access

Conclusions

1. Annotate your maps to have better performance of ., or at least try to express as much code in pattern matching as possible
2. Rewrite nested access like map.submap.subsubmap.field into pattern-matching, or at least use Pathex.
3. Use Tria optimizing compiler which performs analysis to simplify the case generated by ., thus extracting the best performance for the . expression.