defined?(yield) vs block_given?

[davekaro]

Starting in 5.49.0 sequel switched to using "defined?(yield) instead of block_given? internally for better performance on CRuby". I ran some recent benchmarks and found block_given? to be faster. I realized the reason was I had YJIT enabled, and in ruby/ruby#7202 block_given? became YJIT optimized. How do you think about YJIT with respect to performance optimizations like this? Should sequel change again to take advantage of YJIT, but end up being slower if you don't use it? Something else?

[jeremyevans]

Are you testing on Ruby 3.3? Looks like YJIT didn't start optimizing defined?(yield) until after the Ruby 3.3 release. See ruby/ruby@3c92901. I assume with current ruby master, defined?(yield) is faster. It should always be faster, because it doesn't have to check whether the method has been overridden.

[davekaro]

Yes Ruby 3.3. Here is my script:

require "benchmark"
require "benchmark/ips"

RubyVM::YJIT.enable

def defined_yield
  yield if defined?(yield)
end

def block_given
  yield if block_given?
end

Benchmark.ips do |x|
  x.report("defined?(yield)") {
    defined_yield
  }

  x.report("block_given?") {
    block_given
  }

  x.compare!
end

and the results:

ruby bench.rb
ruby 3.3.0 (2023-12-25 revision 5124f9ac75) +YJIT [arm64-darwin23]
Warming up --------------------------------------
     defined?(yield)     3.268M i/100ms
        block_given?     4.010M i/100ms
Calculating -------------------------------------
     defined?(yield)     59.950M (± 0.2%) i/s -    300.674M in   5.015402s
        block_given?     64.555M (± 1.1%) i/s -    324.842M in   5.032632s

Comparison:
        block_given?: 64555304.3 i/s
     defined?(yield): 59950475.8 i/s - 1.08x  slower

If I remove the RubyVM::YJIT.enable this is my result

ruby bench.rb
ruby 3.3.0 (2023-12-25 revision 5124f9ac75) [arm64-darwin23]
Warming up --------------------------------------
     defined?(yield)     3.138M i/100ms
        block_given?     2.090M i/100ms
Calculating -------------------------------------
     defined?(yield)     31.553M (± 1.2%) i/s -    160.013M in   5.072006s
        block_given?     21.130M (± 2.5%) i/s -    106.611M in   5.048559s

Comparison:
     defined?(yield): 31552684.4 i/s
        block_given?: 21130127.9 i/s - 1.49x  slower

Looks like Ruby 3.3 is optimizing it.

[davekaro]

Sorry @jeremyevans after I typed that all out, realized you said defined?(yield) wasn't optimized yet! Your reasoning does make sense to me. Thanks!