Issue #19693 has been updated by byroot (Jean Boussier). Status changed from Open to Closed Should we close? I don't see anything actionable here. ---------------------------------------- Misc #19693: Data initialization is significantly slower than Struct https://bugs.ruby-lang.org/issues/19693#change-103304 * Author: janosch-x (Janosch Müller) * Status: Closed * Priority: Normal ---------------------------------------- Maybe there is potential to make it as fast as Struct? ```ruby require 'benchmark/ips' S = Struct.new(:a, :b, :c, :d, :e, :f, :g, :h, :i, :j) D = Data.define(:a, :b, :c, :d, :e, :f, :g, :h, :i, :j) Benchmark.ips do |x| x.report('Struct') { S.new(1, 2, 3, 4, 5, 6, 7, 8, 9, 10) } x.report('Data') { D.new(1, 2, 3, 4, 5, 6, 7, 8, 9, 10) } x.compare! end; 1 # => [...] # => Struct: 6916530.4 i/s # => Data: 1507259.5 i/s - 4.59x slower ``` -- https://bugs.ruby-lang.org/

Issue #19693 has been updated by byroot (Jean Boussier). @janosch-x based on the discussion in #19278, I don't think it's possible as `Data` specifically designed so that `initialize` always receive keyword arguments. So given the spec (which we can't change) I don't see how it could be made performant, but maybe someone has an idea? ---------------------------------------- Misc #19693: Data initialization is significantly slower than Struct https://bugs.ruby-lang.org/issues/19693#change-103309 * Author: janosch-x (Janosch Müller) * Status: Closed * Priority: Normal ---------------------------------------- Maybe there is potential to make it as fast as Struct? ```ruby require 'benchmark/ips' S = Struct.new(:a, :b, :c, :d, :e, :f, :g, :h, :i, :j) D = Data.define(:a, :b, :c, :d, :e, :f, :g, :h, :i, :j) Benchmark.ips do |x| x.report('Struct') { S.new(1, 2, 3, 4, 5, 6, 7, 8, 9, 10) } x.report('Data') { D.new(1, 2, 3, 4, 5, 6, 7, 8, 9, 10) } x.compare! end; 1 # => [...] # => Struct: 6916530.4 i/s # => Data: 1507259.5 i/s - 4.59x slower ``` -- https://bugs.ruby-lang.org/

Issue #19693 has been updated by nobu (Nobuyoshi Nakada). This is a kind of cheat, but... https://github.com/nobu/ruby/tree/data-optimize --- compare-ruby: ruby 3.3.0dev (2023-05-26T17:07:47Z master c6e4337a99) [x86_64-darwin22] built-ruby: ruby 3.3.0dev (2023-05-27T10:07:26Z data-optimize 81baba152b) [x86_64-darwin22] last_commit=[Misc #19693] Optimize Data.new warming up.... ##### Iteration per second (i/s) | |compare-ruby|built-ruby| |:------------|-----------:|---------:| |Struct-list | 7.194M| 7.169M| | | 1.00x| -| |Struct-kwd | 933.437k| 925.462k| | | 1.01x| -| |Data-list | 921.171k| 9.232M| | | -| 10.02x| |Data-kwd | 944.031k| 946.684k| | | -| 1.00x| ---------------------------------------- Misc #19693: Data initialization is significantly slower than Struct https://bugs.ruby-lang.org/issues/19693#change-103327 * Author: janosch-x (Janosch Müller) * Status: Closed * Priority: Normal ---------------------------------------- Maybe there is potential to make it as fast as Struct? ```ruby require 'benchmark/ips' S = Struct.new(:a, :b, :c, :d, :e, :f, :g, :h, :i, :j) D = Data.define(:a, :b, :c, :d, :e, :f, :g, :h, :i, :j) Benchmark.ips do |x| x.report('Struct') { S.new(1, 2, 3, 4, 5, 6, 7, 8, 9, 10) } x.report('Data') { D.new(1, 2, 3, 4, 5, 6, 7, 8, 9, 10) } x.compare! end; 1 # => [...] # => Struct: 6916530.4 i/s # => Data: 1507259.5 i/s - 4.59x slower ``` -- https://bugs.ruby-lang.org/

Issue #19693 has been updated by janosch-x (Janosch Müller).

I suppose it wouldn't hurt.

I think its beautiful 😁 Another option could be to bypass the initializer by default and only patch it into a Data subclass when `initialize` is redefined (as detected via `method_added` or so).

Defining a `new` singleton method on the Data subclass, in Ruby and with explicit keyword arguments should be able to then use the literal kwargs optimization

@Eregon Wouldn't that require exposing some kind of alternate, non-freezing initializer to Ruby, or some setter methods? This seems to go a bit against the immutable spirit of the class. Also, wouldn't defining `new` in Ruby slow down initialization with positional/list arguments? Maybe the solution would rather be to auto-define a `keyword_initialize` method or so in Ruby and let `new`, still in C, delegate to that method when given kwargs? BTW using literal kwargs in Ruby also makes struct initialization with keywords 3 times faster: ```ruby require 'benchmark/ips' S1 = Struct.new(:a, :b, :c, :d, :e, :f, :g, :h, :i, :j) S2 = Struct.new(:a, :b, :c, :d, :e, :f, :g, :h, :i, :j) class << S2 alias orig_new new def new(a: nil, b: nil, c: nil, d: nil, e: nil, f: nil, g: nil, h: nil, i: nil, j: nil) orig_new(a, b, c, d, e, f, g, h, i, j) end end Benchmark.ips do |x| x.report('ckw') { S1.new(a: 1, b: 2, c: 3, d: 4, e: 5, f: 6, g: 7, h: 8, i: 9, j: 10) } x.report('rubykw') { S2.new(a: 1, b: 2, c: 3, d: 4, e: 5, f: 6, g: 7, h: 8, i: 9, j: 10) } x.compare! end; 1 # rubykw: 4510892.9 i/s # ckw: 1377423.3 i/s - 3.27x slower ``` ---------------------------------------- Misc #19693: Data initialization is significantly slower than Struct https://bugs.ruby-lang.org/issues/19693#change-103331 * Author: janosch-x (Janosch Müller) * Status: Closed * Priority: Normal ---------------------------------------- Maybe there is potential to make it as fast as Struct? ```ruby require 'benchmark/ips' S = Struct.new(:a, :b, :c, :d, :e, :f, :g, :h, :i, :j) D = Data.define(:a, :b, :c, :d, :e, :f, :g, :h, :i, :j) Benchmark.ips do |x| x.report('Struct') { S.new(1, 2, 3, 4, 5, 6, 7, 8, 9, 10) } x.report('Data') { D.new(1, 2, 3, 4, 5, 6, 7, 8, 9, 10) } x.compare! end; 1 # => [...] # => Struct: 6916530.4 i/s # => Data: 1507259.5 i/s - 4.59x slower ``` -- https://bugs.ruby-lang.org/

Issue #19693 has been updated by janosch-x (Janosch Müller). @Eregon Just in case you are interested, I tried implementing your suggestion:

Defining a `new` singleton method on the Data subclass, in Ruby and with explicit keyword arguments should be able to then use the literal kwargs optimization

Unfortunately it doesn't seem to be possible in a way that is performant and spec-compliant. Struct and Data support both list and keyword arguments in `new` now, so an override in Ruby needs to accept both types of arguments, pick one type, and check it for completeness. I [tried to get around checking by using the anonymous kwrest splat](https://gist.github.com/jaynetics/386dda6a576edd942291866c78a0db5f) (`**`) and passing that on to another generated method with explicit kwargs, but `**` brings the performance back to the original level. I guess it also (pointlessly?) creates a Hash. ---------------------------------------- Misc #19693: Data initialization is significantly slower than Struct https://bugs.ruby-lang.org/issues/19693#change-103351 * Author: janosch-x (Janosch Müller) * Status: Closed * Priority: Normal ---------------------------------------- Maybe there is potential to make it as fast as Struct? ```ruby require 'benchmark/ips' S = Struct.new(:a, :b, :c, :d, :e, :f, :g, :h, :i, :j) D = Data.define(:a, :b, :c, :d, :e, :f, :g, :h, :i, :j) Benchmark.ips do |x| x.report('Struct') { S.new(1, 2, 3, 4, 5, 6, 7, 8, 9, 10) } x.report('Data') { D.new(1, 2, 3, 4, 5, 6, 7, 8, 9, 10) } x.compare! end; 1 # => [...] # => Struct: 6916530.4 i/s # => Data: 1507259.5 i/s - 4.59x slower ``` -- https://bugs.ruby-lang.org/

Issue #19693 has been updated by janosch-x (Janosch Müller). @byroot nice idea! i've [tried it](https://gist.github.com/jaynetics/0d04eb85177140e2a84fc88aeea68764). it still doubles (instead of triples) kw init speed, at the cost of halving speed when initializing with list args: ``` orig list: 7136048.0 i/s patch list: 3115468.7 i/s - 2.29x slower patch kw: 2956280.5 i/s - 2.41x slower orig kw: 1621000.3 i/s - 4.40x slower ```

I'm not sure is spec compliant, as it would allow to call Foo.new("a", a: "also_a")

i think this one test, for structs with `keyword_init: true`, will fail: https://github.com/ruby/ruby/blame/master/test/ruby/test_struct.rb#L116 however, this gives me the idea that a simple Ruby keyword initializer (`def self.new(a: nil, b: nil); orig_new(a, b) end`) could be auto-defined, and speed tripled, if and only if `keyword_init: true` is given. this should be spec compliant. structs with `keyword_init: true` already raise an arity error when given list args. then again, i assume the idea is to rather fade out the `keyword_init` option, which doesn't even exist for `Data`? ---------------------------------------- Misc #19693: Data initialization is significantly slower than Struct https://bugs.ruby-lang.org/issues/19693#change-103397 * Author: janosch-x (Janosch Müller) * Status: Closed * Priority: Normal ---------------------------------------- Maybe there is potential to make it as fast as Struct? ```ruby require 'benchmark/ips' S = Struct.new(:a, :b, :c, :d, :e, :f, :g, :h, :i, :j) D = Data.define(:a, :b, :c, :d, :e, :f, :g, :h, :i, :j) Benchmark.ips do |x| x.report('Struct') { S.new(1, 2, 3, 4, 5, 6, 7, 8, 9, 10) } x.report('Data') { D.new(1, 2, 3, 4, 5, 6, 7, 8, 9, 10) } x.compare! end; 1 # => [...] # => Struct: 6916530.4 i/s # => Data: 1507259.5 i/s - 4.59x slower ``` -- https://bugs.ruby-lang.org/