Environment

Rubies:

• jruby 9.1.8.0 (2.3.1) 2017-03-06 90fc7ab Java HotSpot(TM) 64-Bit Server VM 25.131-b11 on 1.8.0_131-b11 +jit [linux-x86_64]

Linux: Linux maruchan 4.10.0-20-generic #22-Ubuntu SMP Thu Apr 20 09:22:42 UTC 2017 x86_64 x86_64 x86_64 GNU/Linux

Expected Behavior

Performance on the benchmark below consistently regresses when running JRuby with -Xcompile.invokedynamic=true.

I was under the impression that after warmup, running JRuby with invokedynamic should never be slower than without it, but that does not seem to be the case on this benchmark.

require 'benchmark/ips'

require 'oat'
require 'oat/adapters/hal'

class Foo
  attr_reader :foo, :bar, :baz

  def initialize(**values)
    values.each do |(key, value)|
      respond_to?(key) ? instance_variable_set(:"@#{key}", value) : raise
    end
  end
end

class FooSerializer < Oat::Serializer
  adapter Oat::Adapters::HAL

  schema do
    property :foo, item.foo
    property :bar, item.bar
    property :baz, item.baz
  end
end

puts "Running with #{RUBY_DESCRIPTION} (JRUBY_OPTS: #{ENV['JRUBY_OPTS']})"

FOO = Foo.new(foo: 1, bar: 2, baz: 'baz')

Benchmark.ips do |benchmark|
  benchmark.time = 120
  benchmark.warmup = 120

  benchmark.report('FooSerializer') { FooSerializer.new(FOO).to_hash }

  benchmark.compare!
end

Actual Behavior

• Without invokedynamic: FooSerializer 251.627k (±23.9%) i/s - 28.264M in 120.011879s
• With invokedynamic: FooSerializer 33.785k (± 6.9%) i/s - 4.032M in 120.030513s

I've also tested with latest master and while the numbers go up slightly in either case, the general relation between the two results keeps the same.