Issue #15836 has been updated by byroot (Jean Boussier).


@mame re benchmark

So I decided to run this against redmine boot, using this branch: https://github.com/redmine/redmine/compare/master...byroot:boot-benchmark

Eager loading is enabled so that the entire codebase is loaded, and it uses https://github.com/SamSaffron/memory_profiler to measure allocations and retentions.

Full benchmark output: https://gist.github.com/byroot/845a5877c1cde91c50b43be446dfb20f

Baseline (official 2.6.3):

```
Total allocated: 121.11 MB (1234362 objects)
Total retained:  24.86 MB (200539 objects)

allocated memory by class
-----------------------------------
  63.36 MB  String

allocated objects by class
-----------------------------------
    980623  String
```

With the patch (official 2.6.3 + this patch):

```
Total allocated: 120.01 MB (1206699 objects)
Total retained:  24.82 MB (199397 objects)

allocated memory by class
-----------------------------------
  62.25 MB  String


allocated objects by class
-----------------------------------
    952953  String
```

Diff:

```
-27 663 allocations (-2.24%)
-1.10MB allocations (-0.9%)
-1 142 retentions
-0.4MB retentions  
```

IMHO that is significant, especially for a small sized application like Redmine. However I can't say wether it outweigh the backward compatibility concern or not. 

### Backward compatibility

One thing to note is that I had to patch https://github.com/rails/rails/blob/28aca474d48b6acdbe8c7861d9347e27c65fafd9/activesupport/lib/active_support/ordered_options.rb#L43 because it was mutating the result of `Symbol#to_s`. Also running the Redmine test suite shows a couple breakage in the i18n gem.

IMHO these are fairly simple to fix, but I would totally understand if that was considered as a no-go.

### Typical code benefiting from this change

  - Rails autoloader and Zeitwerk would both benefit from the `Module#name` change as they both keep references to class names as hash keys
  - Various parts of Rails would benefit as well since they use the class names extensively to derive other class names, as well as symbols `belongs_to :post`.
  - `def method_missing` very often call `name.to_s` to match the method name, hence would benefit from the `Symbol#to_s` as well.
  - Serialization of symbols into various formats, e.g. `{foo: 42}.to_json`. That pattern is fairly common IMO.

----------------------------------------
Feature #15836: [Proposal] Make Module#name and Symbol#to_s return their internal fstrings
https://bugs.ruby-lang.org/issues/15836#change-77957

* Author: byroot (Jean Boussier)
* Status: Feedback
* Priority: Normal
* Assignee: 
* Target version: 
----------------------------------------
# Why ?

In many codebases, especially Rails apps, these two methods are the source of quite a lot of object allocations.

`Module#name` is often accessed for various introspection features, autoloading etc.

`Symbol#to_s` is access a lot by HashWithIndifferentAccess other various APIs accepting both symbols and strings. 

Returning fstrings for both of these methods could significantly reduce allocations, as well as sligthly reduce retention as it would reduce some duplications.

Also, more and more Ruby APIs are now returning fstrings. `frozen_string_literal`AFAIK should become the default some day, string used as hash keys are now automatically interned as well.

### Backward compatibilty 

Of course this is not fully backward compatible, it's inevitable that some code in the wild is mutating the strings returned by these methods, but I do believe it's a rare occurence, and easy to fix. 

### Implementation

I implemented it here: https://github.com/ruby/ruby/pull/2175



-- 
https://bugs.ruby-lang.org/

Unsubscribe: <mailto:ruby-core-request / ruby-lang.org?subject=unsubscribe>
<http://lists.ruby-lang.org/cgi-bin/mailman/options/ruby-core>