Feature #19355: Add finer coarsed locking for searching callable method entries

• Status: Open
• Assignee: -
• Target version: -

[luke-gru (Luke Gruber)]

During multi-ractor mode, unfortunately when calling methods often there's a call to the internal cruby function callable_method_entry, which
returns a cached CME or a new one. In the case that it's cached, there's a VM lock around its access that could be reduced to a lock
pertaining to the class of the caller of the method. This reduces locking in multi-ractor mode and allows, for example, object allocation to happen
at the same time as method calls.

[luke-gru (Luke Gruber)]

MyClass = Object
class MyClass2 < Object;
  def call_a_method
    []
  end
  def call_another_method
    {}
  end
end
USE_RACTORS = ARGV.first == "ractor"
puts "ractors: #{USE_RACTORS ? 't' : 'f'}"

class FakeRactor
  def initialize
    yield
    self
  end
  def take
  end
end

def concurrency(&blk)
  if USE_RACTORS
    Ractor.new(&blk)
  else
    FakeRactor.new(&blk)
  end
end

bench = Benchmark.measure do
  rs = []
  100.times do
    rs << concurrency do
      50_000.times do|i|
        obj = MyClass.new
        obj.respond_to?(:hi)
        obj2 = MyClass2.new
        obj2.respond_to?(:hi)
        obj2.call_a_method
        obj2.call_another_method
      end
    end
  end
  rs.each(&:take)
end
puts bench

On my machine without my local patch, it takes 6.5 seconds and with the patch 5.5 seconds.
Of course storing a mutex on every class instance takes more memory so it might not be worth
it But with this new mutex it's possible also to do constant lookup / update without the VM
lock too.