ZJIT: Something is wrong with types::Class in the HIR type system

[tekknolagi]

Currently breaking rubocop: ruby#14945 (comment)

[tekknolagi]

Looking it's like .class

[tekknolagi]

bisect points to

initialize@/Users/emacs/.gem/ruby/ruby-zjit-rel-stats/gems/parser-3.3.9.0/lib/parser/lexer-strings.rb

[tekknolagi]

Could also be a weird intersection between .class and .respond_to? inlining

[tekknolagi]

Minimal repro:

# miniruby --zjit thisfile.rb
class C
  class << self
    attr_accessor :_lex_actions
    private :_lex_actions, :_lex_actions=
  end
  self._lex_actions = [1, 2, 3]
  def initialize
    @_lex_actions =
      if self.class.respond_to?(:_lex_actions, true)
        self.class.send :_lex_actions
      else
        []
      end
  end
  def num = @_lex_actions.size
end
def assert_equal a, b
  raise "Assertion failed: #{a.inspect} != #{b.inspect}" unless a == b
end
35.times do
  assert_equal C.new.num, 3
end

[tekknolagi]

This is a subtle type system thing, I think

[k0kubun]

If it takes time to fix it correctly, could we consider reverting ruby#14945 or part of it in the mean time? A failing headline benchmark makes it harder to evaluate performance-sensitive changes and detect/bisect other issues on the benchmark.

[tekknolagi]

I think we can instead revert the type::Class annotation for .class

[k0kubun]

works for me 👍

Generally, I want us to treat ruby-bench like a secondary ZJIT CI. When we notice a merged change failed a benchmark, we should instantly fix or revert it. As long as it runs the failed benchmark correctly, any kind of revert works as a fix to it.

[tekknolagi]

We should run ruby-bench on CI

[k0kubun]

We used to run ruby-bench's railsbench with --once on ruby/ruby CI actually. However, even while it was just railsbench, it was unstable enough as a CRuby CI due to bundle install or something that we marked it continue-on-error: true. But we can try again.