[ruby-core:121554] [Ruby Bug#21201] Performance regression when defining methods inside `refine` blocks

Issue #21201 has been updated by byroot (Jean Boussier).


Your patch look really good.

I wonder if it would be possible to do like the `vm->constant_cache` table, have the key be the method name, so that you wouldn't need to clear absolutely everything.

But your patch as-is is already a nice improvement I think, the only downside being the one extra lock point for ractors.

----------------------------------------
Bug #21201: Performance regression when defining methods inside `refine` blocks
https://bugs.ruby-lang.org/issues/21201#change-112581

* Author: alpaca-tc (Hiroyuki Ishii)
* Status: Open
* ruby -v: 3.3.7, 3.4.2, 3.5.0
* Backport: 3.1: UNKNOWN, 3.2: UNKNOWN, 3.3: UNKNOWN, 3.4: UNKNOWN
----------------------------------------
After the change introduced in [PR #10037](https://github.com/ruby/ruby/pull/10037), a significant performance regression has been observed when defining methods within refine blocks.

The PR correctly fixes a bug where `callcache` invalidation was missing when methods are defined inside a refinement. To fix this, it now invokes [`rb_clear_all_refinement_method_cache()`](https://github.com/ruby/ruby/blob/752a1d785475d1950b33c7d77b289a6a3a753f02/vm_method.c#L439-L443) at the time of method definition within `refine`. However, this function performs a full object space traversal via `rb_objspace_each_objects`, making it extremely expensive.

In our team's Rails application, refinements are used heavily. During application code reload (triggered in development mode), `rb_clear_all_refinement_method_cache()` is called 1425 times, 1142 of which originate via `rb_clear_method_cache()`. As a result, the code reload time has increased from **5 seconds** to **15 seconds** after the patch. Since reloading occurs every time the code changes, this degrades development experience severely.

### Reproduction script:
```ruby
require "bundler/inline"

gemfile(true) do
  source "https://rubygems.org"

  gem "benchmark-ips"
end

mod = Module.new
klass = Class.new

Benchmark.ips do |x|
  x.report(RUBY_VERSION) do
    mod.send(:refine, klass) do
      def call_1 = nil
      def call_2 = nil
      def call_3 = nil
    end
  end

  x.save! "/tmp/performance_regression_refine.bench"
  x.compare!
end
```

### Benchmark results:
```
Comparison:
               3.2.7:  1500316.7 i/s
               3.3.7:      158.0 i/s - 9496.46x  slower
               3.5.0:      124.6 i/s - 12041.43x  slower
               3.4.2:      119.5 i/s - 12553.50x  slower
```

Related Slack discussion: https://ruby-jp.slack.com/archives/CLWSHA76V/p1741932018811539  
I totally understand the necessity and value of the past PR that fixed the callcache bugs. That said, I was wondering if there might be any ideas to make it faster, or to handle the invalidation more efficiently.



-- 
https://bugs.ruby-lang.org/
 ______________________________________________
 ruby-core mailing list -- [e-mail]
 To unsubscribe send an email to [e-mail]
 ruby-core info -- https://ml.ruby-lang.org/mailman3/lists/ruby-core.ml.ruby-lang.org/