Assert scope is not match/guard when using escaped regexes

[sabiwara]

Following on #14720, I've been thinking if we could/should improve the error message when trying to use regex module attributes in guards/match scope.
This PR explores the idea, but feel free to close if this is overthinking on my part.

Before (confusing: :re.compile mentioned but not in the code)

After

[josevalim]

I think in this case some code duplication is fine. Let's just reimplement the error message directly in the Regex module, since that's what we would expect library authors to do too.

[josevalim]

Instead of generating additional code, we should call Regex.__import__(unquote(...)), where __import__ itself is a macro which also checks the environment.

[sabiwara]

This sounds much better indeed 👍

[sabiwara]

Arg, this might actually not work, since unlike Kernel we need to require Regex.

       quote do
            require Regex
            Regex.__import_pattern__(unquote(Macro.escape(exported)))
          end

Which then makes it impossible to compile elixir:

==> elixir (compile)
error: you are trying to use/import/require the module Regex which is currently being defined.

(although it seems weird to warn here, since it's happening within quote 🤔)

[sabiwara]

Pushed the broken commit just in case: c1dee89

[josevalim]

I believe the issue happens because we escape the struct defined by defstruct inside Regex. An easy fix is to add this clause to the top of the cond in __escape__:

        is_nil(regex.re_pattern) ->
          nil

And then it compiles fine!

[sabiwara]

Another idea to keep the emitted AST small 💡

We could leave a special meta when we escape a struct with __escape__, and add an extra hint when we emit the scope error if the meta is there? (Hint: This happened because you tried to escape a Regex).

Another benefit is that custom structs need to do nothing to benefit from this.

[josevalim]

Do you think that metadata would work? The __escape__ function can return any node and it would be hard to know which node to annotate/which node would fail.

Another option is for us to allow this to work:

require Regex

quote do
  Regex.__import__(...)
end

or even just this:

quote do
  Regex.__import__(...)
end

Especially as I thought we already annotated within quote that the module was already required...

[sabiwara]

Do you think that metadata would work? The escape function can return any node and it would be hard to know which node to annotate/which node would fail.

That's true, I didn't think this through, sorry.

or even just this:

quote do
  Regex.__import__(...)
end

That would be ideal, but isn't there a risk of allowing implicit requires?
Wouldn't this mess with the lexical tracker?

[josevalim]

Ah yes, the way it works is that we annotate it is required and consider the require in both caller and callee. At least that’s how we handle imports and they are more complex than require.

another option is to allow requiring the current module as long as it happens inside a function? As you can assume that when the function is executed the module is indeed defined.

[sabiwara]

TIL about the required: true meta, I think it fixed it nicely 0e671f4.

[sabiwara]

OK to backport?

[josevalim]

@sabiwara yes, let's backport!