Unhandled format in elixir_erl_compiler:format_errors/1 throws argument error

[marianoguerra]

Precheck

• Do not use the issue tracker for help or support (try Elixir Forum, Stack Overflow, IRC, etc.)
• For proposing a new feature, please start a discussion on the Elixir Core mailing list: https://groups.google.com/group/elixir-lang-core
• For bugs, do a quick search and make sure the bug has not yet been reported
• Please disclose security vulnerabilities privately at elixir-security@googlegroups.com
• Finally, be nice and have fun!

Environment

• Elixir & Erlang/OTP versions (elixir --version):

Erlang/OTP 23 [erts-11.1] [source] [64-bit] [smp:8:8] [ds:8:8:10] [async-threads:1] [hipe]
Elixir 1.10.4 (compiled with Erlang/OTP 23)

• Operating system: Ubuntu 20.04

Current behavior

Include code samples, errors and stacktraces if appropriate.

When compiling the following code with elixirc mymod.ex

defmodule MyMod do
  def f(r) when :erlang.is_record(r, :my_record) do
    :ok
  end
end

This error is printed:

== Compilation error in file mymod.ex ==
** (ArgumentError) argument error
    (stdlib 3.13.2) :unicode.characters_to_binary(MyMod)
    (elixir 1.10.4) src/elixir_utils.erl:125: :elixir_utils.characters_to_binary/1
    (elixir 1.10.4) src/elixir_erl_compiler.erl:76: anonymous fn/1 in :elixir_erl_compiler.format_errors/1
    (stdlib 3.13.2) lists.erl:1342: :lists.foreach/2
    (elixir 1.10.4) src/elixir_erl_compiler.erl:12: anonymous fn/3 in :elixir_erl_compiler.spawn/2

Tracing the call:

:dbg.start()
:dbg.tracer()
:dbg.p(:all, :c)
:dbg.tpl(:elixir_erl_compiler, :format_errors, 1, [{:"_", [], [{:return_trace}]}])
Elixir.Kernel.ParallelCompiler.compile(["mymod.ex"])

Shows this format:

elixir_erl_compiler:format_errors([{'Elixir.MyMod',[{none,core_lint,{illegal_guard,{f,1}}}]}])

Expected behavior

Either the elixir compiler rejecting the guard if it makes sense or changing

elixir/lib/elixir/src/elixir_erl_compiler.erl

Line 76 in 3415f90

BinFile = elixir_utils:characters_to_binary(File),

To something like Source = format_error_source(File) that handles atoms.

[michalmuskala]

This is probably because Elixir bypasses some passes of the Erlang compiler and is_record/2 is not an actual guard function, but a "magic" one that is expanded by the compiler to is_record/3 in one of those skipped passes. I believe Elixir should just refuse to compile this with a nice error message, given it has a different mechanism for records.

[josevalim]

We validate if something is a valid guard but is_record is passing the check for some reason, probably because we use erl_internal. It should be an easy fix, thanks!

[michalmuskala]

I think the fix is wrong - is_record/3 should be allowed - it's basically an alias to is_tuple(x) and tuple_size(x) == size and element(1, x) == tag. The is_record/2 is problematic since it's a special function that relies on -record definitions, but the /3 variant is a regular function.

[josevalim]

is_record is a compiler macro though? Or is it an actual function? If the former, then we don’t implement it, so the fix is still good. :)

[michalmuskala]

is_record/2 is macro-like, is_record/3 is a regular guard

[josevalim]

@michalmuskala the core erlang pass is still unhappy when I pass erlang:is_record/3 as is.

[josevalim]

@michalmuskala oh, it expects the second and third arguments to be literals. I am thinking it may be best to continue forbidding both of them after all, as I would like to avoid special casing validation around :erlang functions in our compiler. 🤔

[michalmuskala]

Ah, you're right. I did not realise it requires those to be literals in a guard.

I still think disallowing it could be a breaking change, though.

[josevalim]

Given the function was never available without the :erlang prefix, :erlang.is_record/2 was on v1.9, and none of them already doesn't work on v1.11, the odds are that nobody actually uses it. Given it doesn't work on v1.11, we will keep the good error message for now and revert only if anybody reports a regression based on v1.11.

Thank you!