Autocorrecting a Style/RedundantSelfAssignmentBranch offence breaks existing code

[hugovandevliert]

While upgrading our codebase to RuboCop 1.19.0 and running bundle exec rubocop -a, I noticed some failing specs. I believe the Style/RedundantSelfAssignmentBranch cop autocorrections should be marked unsafe. See the example below.

---

Expected behavior

bundle exec rubocop -a should not break existing code.

Actual behavior

bundle exec rubocop -a broke some existing code.

Steps to reproduce the problem

Let's say we have the following (rather odd) piece of code in test.rb:

class Test
  def foo
    @foo = if @foo_set
      @foo
    else
      @foo_set = true
      @foo = 'bar'
    end
  end
end

test = Test.new
puts test.foo == 'bar' # should be true
puts test.foo == 'bar' # should be true

Running this file will give the following output:

true
true

RuboCop will display the following message on line 4:

Style/RedundantSelfAssignmentBranch: Remove the self-assignment branch.

Executing bundle exec rubocop -a test.rb will update the code to this:

class Test
  def foo
    @foo = unless @foo_set
      @foo_set = true
      @foo = 'bar'
    end
  end
end

test = Test.new
puts test.foo == 'bar' # should be true
puts test.foo == 'bar' # should be true

Running this file will give a different output:

true
false

RuboCop version

1.19.0 (using Parser 3.0.2.0, rubocop-ast 1.10.0, running on ruby 2.7.4 x86_64-darwin20)
  - rubocop-performance 1.11.4
  - rubocop-rails 2.11.3
  - rubocop-rspec 2.4.0