ERL-1179: Compiler/runtime yields invalid semantic behaviour (boolean value is flipped)

[OTP-Maintainer]

Original reporter: josevalim
Affected version: OTP-23
Fixed in version: OTP-23.0
Component: compiler
Migrated from: https://bugs.erlang.org/browse/ERL-1179

---

This bug was found in master at commit:
{code:java}
commit a4e8656d1aa4bde0bbfab6a52585add5d7951bd1
Merge: b7d1866d48 75d4701fff
Author: Sverker Eriksson <sverker@erlang.org>
Date:   Tue Feb 18 16:38:24 2020 +0100
{code}
Take the following Erlang code:
{code:java}
-module(foo).
-export([bar/2]).

bar(_, Caller) ->
  erlang:display({find(module, Caller), find(function, Caller), find(context, Caller)}),
  IsFunction = find(function, Caller) /= nil,

  case not true of
    true ->
      ok;
    false ->
      case (not IsFunction) andalso (find(context, Caller) == match) of
        true -> ok;
        false -> erlang:display({IsFunction, Caller})
      end
  end.

-compile({inline, [find/2]}).

find(Key, Map) ->
  case Map of
    #{Key := Value} -> Value;
    #{} -> erlang:error(badmap)
  end.
{code}
Compile it with `erlc foo.erl` and now in `erl` type this:
{code:java}
1> foo:bar(baz, #{module => regex, function => nil, context => nil}).
{regex,nil,nil}
{true,#{context=>nil,function=>nil,module=>regex}}
{code}
As you can see, it is saying that `IsFunction` is true but it should be false (as it is not different to nil).

Thank you!

[OTP-Maintainer]

john said:

Thanks for your report! It's a bit too late to get a fix into OTP 23 RC1, but I've got a preliminary one below:

https://github.com/jhogberg/otp/tree/john/compiler/ssa_opt_ne-flipped/ERL-1179

[OTP-Maintainer]

josevalim said:

Thank you for the fix. Btw, do you have a straight-forward workaround to suggest? Without this fix, Elixir won't be usable with Erlang/OTP 23 RC1. But if there is a simple fix, I can release a new Elixir version with it. :)

[OTP-Maintainer]

john said:

The fix is to not use {{/=}} or {{=/=}}, which might be a bit tricky. :D

Having slept on it I've decided to revert the change that caused this since there's little point to releasing RC1 with such a serious bug present. We'll include a variation of the above fix in RC2.

[OTP-Maintainer]

josevalim said:

Ah, thank you! I have figured out that adding a "_ = atom_to_binary(IsFunction, utf8)" also fixes it but you are right, other code may be affected by this, so reverting the optimization before release makes sense.

[OTP-Maintainer]

bjorn said:

The optimization has been reintroduced in a safe way in master and will be included in RC2.