Inline protocol implementations

[josevalim]

Today, a protocol definition and implementation happens as follow:

defprotocol Binary.Chars do
  def to_binary(thing)
end

defimpl Binary.Chars, for: BitString do
  def to_binary(thing) when is_binary(thing) do
    thing
  end
end

When one calls:

Binary.Chars.to_binary "hello"

It is then dispatched to:

Binary.Chars.BitString.to_binary "hello"

I want to allow the following construct as well:

defprotocol Binary.Chars do
  def to_binary(thing)

  defimpl for: BitString do
    def to_binary(thing) when is_binary(thing) do
      thing
    end
  end
end

The example above will inline the to_binary implementation with the protocol one, giving the following advantages:

1. We skip a remote call for the inlined implementations
2. We avoid generating .beam for a module

However:

1. It means a Binary.Chars.BitString is no longer available, so we lose the property that every protocol implementation contains a module representing it (since in this case the implementation is inlined)

If we are proceeding with this task, we should:

• Provide the inlining implementation
• Set @moduledoc false for all defimpl for default types
• Update exdoc to include @only and @except values on protocol pages
• Deprecate __impl_for__ reflection from protocols (which I don't believe is used)

[yrashk]

wrt to (1), may be we can still generate the module, but do the inlining anyway?

[josevalim]

@yrashk We could but skipping the .beam file is half of the benefit imo. Just the inlining per se is not worth the amount of work imo.

[rafaelfranca]

What are the implication of not generating the module besides the loss of the property?

[yrashk]

How about we make .beam file generation... optional (disabled by default for inlined impls?)

[josevalim]

@rafaelfranca I cannot think of anything we would lose. I raised this issue to see if someone can come up with extra drawbacks, including regarding the beam file. :)

[devinus]

👍

This would allow me to implement Dict with a Protocol and not sacrifice performance.

[josevalim]

Closed in favor #950