-
Notifications
You must be signed in to change notification settings - Fork 15
Optimize DelegateClass using ... delegation
#46
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Conversation
22ee1ad to
5c8ab6c
Compare
Fix: rails#56167 I opened ruby/delegate#46 upstream to improve `ClassDelegator` but it has a bunch of backward compatibility concerns that limits how much it can be optimized. For internal use cases, we can define our own private class delegator and optimize it much further. ``` == no arguments == ruby 3.4.6 (2025-09-16 revision dbd83256b1) +YJIT +PRISM [arm64-darwin24] Warming up -------------------------------------- DelegateClass 1.014M i/100ms Opt 3.273M i/100ms Calculating ------------------------------------- DelegateClass 12.322M (± 0.8%) i/s (81.15 ns/i) - 61.838M in 5.018661s Opt 56.866M (± 0.4%) i/s (17.59 ns/i) - 284.716M in 5.006894s Comparison: DelegateClass: 12322457.6 i/s Opt: 56865721.7 i/s - 4.61x faster == many arguments == ruby 3.4.6 (2025-09-16 revision dbd83256b1) +YJIT +PRISM [arm64-darwin24] Warming up -------------------------------------- DelegateClass 555.419k i/100ms Opt 3.214M i/100ms Calculating ------------------------------------- DelegateClass 6.256M (± 0.6%) i/s (159.85 ns/i) - 31.659M in 5.060874s Opt 53.902M (± 0.0%) i/s (18.55 ns/i) - 269.936M in 5.007857s Comparison: DelegateClass: 6255815.2 i/s Opt: 53902461.2 i/s - 8.62x faster == optional arguments == ruby 3.4.6 (2025-09-16 revision dbd83256b1) +YJIT +PRISM [arm64-darwin24] Warming up -------------------------------------- DelegateClass 826.558k i/100ms Opt 1.645M i/100ms Calculating ------------------------------------- DelegateClass 9.821M (± 0.8%) i/s (101.82 ns/i) - 49.593M in 5.049895s Opt 21.177M (± 0.8%) i/s (47.22 ns/i) - 106.952M in 5.050788s Comparison: DelegateClass: 9821329.7 i/s Opt: 21176752.0 i/s - 2.16x faster ``` Benchmark: ```ruby require "delegate" require "bundler/inline" gemfile do gem "benchmark-ips" gem "rails", path: "." end class User attr_accessor :name def initialize(name) @name = name end def do_something(a, b, c, d:) :something end def opt_args(a=1, b: 2) :something end end StdlibDelegator = DelegateClass(User) ASDelegator = ActiveSupport::Delegation::DelegateClass(User) direct = User.new("George") stdlib = StdlibDelegator.new(direct) opt = ASDelegator.new(direct) puts "== no arguments ==" Benchmark.ips do |x| x.report("DelegateClass") { stdlib.name } x.report("Opt") { opt.name } x.compare!(order: :baseline) end puts "== many arguments ==" Benchmark.ips do |x| x.report("DelegateClass") { stdlib.do_something(1, 2, 3, d: 4) } x.report("Opt") { opt.do_something(1, 2, 3, d: 4) } x.compare!(order: :baseline) end puts "== optional arguments ==" Benchmark.ips do |x| x.report("DelegateClass") { stdlib.opt_args } x.report("Opt") { opt.opt_args } x.compare!(order: :baseline) end ```
Fix: rails#56167 I opened ruby/delegate#46 upstream to improve `ClassDelegator` but it has a bunch of backward compatibility concerns that limits how much it can be optimized. For internal use cases, we can define our own private class delegator and optimize it much further. ``` == no arguments == ruby 3.4.6 (2025-09-16 revision dbd83256b1) +YJIT +PRISM [arm64-darwin24] Warming up -------------------------------------- DelegateClass 1.014M i/100ms Opt 3.273M i/100ms Calculating ------------------------------------- DelegateClass 12.322M (± 0.8%) i/s (81.15 ns/i) - 61.838M in 5.018661s Opt 56.866M (± 0.4%) i/s (17.59 ns/i) - 284.716M in 5.006894s Comparison: DelegateClass: 12322457.6 i/s Opt: 56865721.7 i/s - 4.61x faster == many arguments == ruby 3.4.6 (2025-09-16 revision dbd83256b1) +YJIT +PRISM [arm64-darwin24] Warming up -------------------------------------- DelegateClass 555.419k i/100ms Opt 3.214M i/100ms Calculating ------------------------------------- DelegateClass 6.256M (± 0.6%) i/s (159.85 ns/i) - 31.659M in 5.060874s Opt 53.902M (± 0.0%) i/s (18.55 ns/i) - 269.936M in 5.007857s Comparison: DelegateClass: 6255815.2 i/s Opt: 53902461.2 i/s - 8.62x faster == optional arguments == ruby 3.4.6 (2025-09-16 revision dbd83256b1) +YJIT +PRISM [arm64-darwin24] Warming up -------------------------------------- DelegateClass 826.558k i/100ms Opt 1.645M i/100ms Calculating ------------------------------------- DelegateClass 9.821M (± 0.8%) i/s (101.82 ns/i) - 49.593M in 5.049895s Opt 21.177M (± 0.8%) i/s (47.22 ns/i) - 106.952M in 5.050788s Comparison: DelegateClass: 9821329.7 i/s Opt: 21176752.0 i/s - 2.16x faster ``` Benchmark: ```ruby require "delegate" require "bundler/inline" gemfile do gem "benchmark-ips" gem "rails", path: "." end class User attr_accessor :name def initialize(name) @name = name end def do_something(a, b, c, d:) :something end def opt_args(a=1, b: 2) :something end end StdlibDelegator = DelegateClass(User) ASDelegator = ActiveSupport::Delegation::DelegateClass(User) direct = User.new("George") stdlib = StdlibDelegator.new(direct) opt = ASDelegator.new(direct) puts "== no arguments ==" Benchmark.ips do |x| x.report("DelegateClass") { stdlib.name } x.report("Opt") { opt.name } x.compare!(order: :baseline) end puts "== many arguments ==" Benchmark.ips do |x| x.report("DelegateClass") { stdlib.do_something(1, 2, 3, d: 4) } x.report("Opt") { opt.do_something(1, 2, 3, d: 4) } x.compare!(order: :baseline) end puts "== optional arguments ==" Benchmark.ips do |x| x.report("DelegateClass") { stdlib.opt_args } x.report("Opt") { opt.opt_args } x.compare!(order: :baseline) end ```
lib/delegate.rb
Outdated
| public_instance_methods -= ignores | ||
|
|
||
| instance_methods = (public_instance_methods + protected_instance_methods) | ||
| normal, special = instance_methods.partition { |m| m.match?(/\A[a-zA-Z]\w*!?\z/) } |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
| normal, special = instance_methods.partition { |m| m.match?(/\A[a-zA-Z]\w*!?\z/) } | |
| normal, special = instance_methods.partition { |m| m.match?(/\A[a-zA-Z]\w*[!?]?\z/) } |
Il also saw that in the rails patch you used \w+ rather than \w*, IDK if there is a good reason for that
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
in the rails patch you used
\w+rather than\w*, IDK if there is a good reason for that
It's a mistake, \w* is correct.
By generating source code for methods that use `...` delegation when possible, we can lower the overhead of delegation by half. This could be lowered further by copying the delegated method signature, like in ruby#16, but this would assume the delegated method signature never change, so I'm not sure if that's OK. Then most of the remaining overhead is in calling `__getobj__`, but that's part of the spec, so can't be eliminated. Results: ``` == no arguments == ruby 3.4.6 (2025-09-16 revision dbd83256b1) +YJIT +PRISM [arm64-darwin24] Warming up -------------------------------------- baseline 3.838M i/100ms handrolled 3.465M i/100ms DelegateClass 979.160k i/100ms Opt 2.028M i/100ms Calculating ------------------------------------- baseline 64.296M (± 0.5%) i/s (15.55 ns/i) - 322.355M in 5.013724s handrolled 57.058M (± 0.4%) i/s (17.53 ns/i) - 287.567M in 5.039966s DelegateClass 12.118M (± 0.5%) i/s (82.52 ns/i) - 60.708M in 5.009812s Opt 27.764M (± 0.5%) i/s (36.02 ns/i) - 139.925M in 5.039997s Comparison: baseline: 64296345.8 i/s handrolled: 57058063.8 i/s - 1.13x slower Opt: 27763713.5 i/s - 2.32x slower DelegateClass: 12118085.0 i/s - 5.31x slower == many arguments == ruby 3.4.6 (2025-09-16 revision dbd83256b1) +YJIT +PRISM [arm64-darwin24] Warming up -------------------------------------- baseline 3.605M i/100ms handrolled 3.275M i/100ms DelegateClass 623.030k i/100ms Opt 1.348M i/100ms Calculating ------------------------------------- baseline 63.349M (± 1.6%) i/s (15.79 ns/i) - 317.272M in 5.009667s handrolled 56.277M (± 0.2%) i/s (17.77 ns/i) - 281.623M in 5.004270s DelegateClass 7.079M (± 4.1%) i/s (141.26 ns/i) - 35.513M in 5.026286s Opt 17.953M (± 0.1%) i/s (55.70 ns/i) - 90.345M in 5.032248s Comparison: baseline: 63348844.0 i/s handrolled: 56276912.5 i/s - 1.13x slower Opt: 17953308.5 i/s - 3.53x slower DelegateClass: 7079118.4 i/s - 8.95x slower ``` Benchmark: ```ruby require "delegate" require "bundler/inline" gemfile do gem "benchmark-ips" end class User attr_accessor :name def initialize(name) @name = name end def do_something(a, b, c, d: 1) :something end end class HandrolledDelegator def initialize(user) @user = user end def name @user.name end def do_something(a, b, c, d: 1) @user.do_something(a, b, c, d: 1) end end StdlibDelegator = DelegateClass(User) def OptDelegateClass(superclass, &block) klass = Class.new(Delegator) ignores = [*::Delegator.public_api, :to_s, :inspect, :=~, :!~, :===] protected_instance_methods = superclass.protected_instance_methods protected_instance_methods -= ignores public_instance_methods = superclass.public_instance_methods public_instance_methods -= ignores instance_methods = (public_instance_methods + protected_instance_methods) normal, special = instance_methods.partition { |m| m.match?(/\A[a-zA-Z]\w*\z/) } source = normal.map do |method| "def #{method}(...); __getobj__.#{method}(...); end" end klass.module_eval do def __getobj__ # :nodoc: unless defined?(@delegate_dc_obj) return yield if block_given? __raise__ ::ArgumentError, "not delegated" end @delegate_dc_obj end def __setobj__(obj) # :nodoc: __raise__ ::ArgumentError, "cannot delegate to self" if self.equal?(obj) @delegate_dc_obj = obj end class_eval(source.join(";"), __FILE__, __LINE__) special.each do |method| define_method(method, Delegator.delegating_block(method)) end protected(*protected_instance_methods) end klass.define_singleton_method :public_instance_methods do |all=true| super(all) | superclass.public_instance_methods end klass.define_singleton_method :protected_instance_methods do |all=true| super(all) | superclass.protected_instance_methods end klass.define_singleton_method :instance_methods do |all=true| super(all) | superclass.instance_methods end klass.define_singleton_method :public_instance_method do |name| super(name) rescue NameError raise unless self.public_instance_methods.include?(name) superclass.public_instance_method(name) end klass.define_singleton_method :instance_method do |name| super(name) rescue NameError raise unless self.instance_methods.include?(name) superclass.instance_method(name) end klass.module_eval(&block) if block return klass end OptStdlibDelegator = OptDelegateClass(User) direct = User.new("George") handrolled = HandrolledDelegator.new(direct) stdlib = StdlibDelegator.new(direct) opt_stdlib = OptStdlibDelegator.new(direct) puts "== no arguments ==" Benchmark.ips do |x| x.report("baseline") { direct.name } x.report("handrolled") { handrolled.name } x.report("DelegateClass") { stdlib.name } x.report("Opt") { opt_stdlib.name } x.compare!(order: :baseline) end puts "== many arguments ==" Benchmark.ips do |x| x.report("baseline") { direct.do_something(1, 2, 3, d: 4) } x.report("handrolled") { handrolled.do_something(1, 2, 3, d: 4) } x.report("DelegateClass") { stdlib.do_something(1, 2, 3, d: 4) } x.report("Opt") { opt_stdlib.do_something(1, 2, 3, d: 4) } x.compare!(order: :baseline) end ```
5c8ab6c to
b2293d4
Compare
This reverts commit fc2bd04. Co-authored-by: Jean Boussier <byroot@ruby-lang.org>
This reverts commit fc2bd04. Co-authored-by: Jean Boussier <byroot@ruby-lang.org>
This reverts commit fc2bd04. Co-authored-by: Jean Boussier <byroot@ruby-lang.org>
Reapply "Merge pull request #46 from byroot/use-forward-send"
This reverts commit fc2bd04. Co-authored-by: Jean Boussier <byroot@ruby-lang.org>
By generating source code for methods that use
...delegation when possible, we can lower the overhead of delegation by half.This could be lowered further by copying the delegated method signature, like in #16, but this would assume the delegated method signature never change, so I'm not sure if that's OK.
Then most of the remaining overhead is in calling
__getobj__, but that's part of the spec, so can't be eliminated.Results:
Benchmark: