Reduce memory usage in IEx autocomplete module listing

[dkuku]

Summary

Reduce memory usage in IEx autocomplete module listing. On memory-capped pods, tab completion could trigger OOM kills (exit 137) due to unnecessary allocations.

Two changes:

• Use :erlang.loaded/0 instead of :code.all_loaded/0 and :lists.usort/1 instead of Enum.sort/1 |> Enum.dedup/1. We only need module names for autocomplete, not file paths. :code.all_loaded/0 copies the full beam file path for every loaded module, which is pure waste. :lists.usort/1 combines sort + dedup in a single native pass.

• Use binary_part + :binary.match instead of String.split + Enum.at in match_elixir_modules. Avoids allocating intermediate lists of string fragments for every matching module.

Measurements on a production pod (~200 Elixir modules)

	Memory per tab completion
Before	15,996 KB
After	1,089 KB

14.7x reduction — from 15.6 MB to 1.1 MB per autocomplete call.

Test plan

• make test_iex — 279 tests, 0 failures
• Verified correctness: old and new produce identical results
• Measured on production pod with ~200 loaded Elixir modules

🤖 Generated with Claude Code

[sabiwara]

NIce!
I noticed a difference between the two, don't sure if it is an issue or not:

iex> (:erlang.loaded()  |> :lists.usort()) -- (:code.all_loaded() |> Enum.map(&elem(&1, 0)) |> :lists.usort())
[:elixir_compiler_0, :elixir_compiler_1, :elixir_compiler_2, :elixir_compiler_3]

[dkuku]

I only checked it like that:

iex(add_ons_all@add-ons-5f6694dff9-2sz5f)17> :erlang.loaded() |> length()
6538
iex(add_ons_all@add-ons-5f6694dff9-2sz5f)18> :code.all_loaded() |> length()
6538

But still on the app I'm testing I'm getting an empty list. This is some older elixir version and a release running on kubernetes where this bothers me the most :D

 (:erlang.loaded()  |> :lists.usort()) -- (:code.all_loaded() |> Enum.map(&elem(&1, 0)) |> :lists.usort())
[]

[josevalim]

I think the compiler modules will be fine because are likely not looking up under :elixir_. If it is a problem we can filter them out later!

[josevalim]

💚 💙 💜 💛 ❤️

[josevalim]

Sorry, there may be a bug here. is it guaranteeed that byte_size(mod) - prefix_size is greater than 0?

[sabiwara]

I think the compiler modules will be fine because are likely not looking up under :elixir_. If it is a problem we can filter them out later!

Yes I totally agree, my point was more about the fact these are not returning exactly the same thing, so I wondered if there was a risk of having a wonky behavior.

But worst case scenario, we can always check for :erlang.module_loaded to avoid old code.

[josevalim]

Oh, I see and good call! We can use module_loaded in the future if we run into something, yeah!