From: "byroot (Jean Boussier) via ruby-core" <ruby-core@...>
Date: 2025-05-20T08:18:30+00:00
Subject: [ruby-core:122203] [Ruby Feature#21353] Add shape_id to RBasic under 32 bit

Issue #21353 has been reported by byroot (Jean Boussier).

----------------------------------------
Feature #21353: Add shape_id to RBasic under 32 bit
https://bugs.ruby-lang.org/issues/21353

* Author: byroot (Jean Boussier)
* Status: Open
----------------------------------------
Currently on 64bit systems, for every types, the `shape_id` is stored inside the `RBasic.flags` field, and is 32bit
long.

However, on 32bit systems like i686 and WASM, it is much more complicated.
For `T_OBJECT`, `T_CLASS` and `T_MODULE`, the `shape_id` is stored as part of "user flags" in `FL_USER4-19`,
and for all other types it's stored alongside the instance variable in the `generic_fields_tbl`, which means
a hash lookup is required to access it.

This situation makes a lot of routine noticeably more complicated, with numerous codepath taken only by 32bit systems,
because to avoid doing two hash-lookup per ivar access, the code need a lot of contortions.
You can look for `SHAPE_IN_BASIC_FLAGS` to have an idea of the added complexity.

In addition, it forces us to duplicate some bits of information. For instance `RUBY_FL_FREEZE` is redundant with the
`shape_id`. The shape already record that the object is frozen, and the only reason `RUBY_FL_FREEZE` hasn't been
eliminated is because on 32bits the `shape_id` isn't store inline for some objects.

Similarly, `RUBY_FL_EXIVAR` is redundant with the `shape_id`, because to know whether an object has ivars, you can
simply check if `shape_id == 0`.

Reclaiming these bits would be very useful for Ractors, as we'd need two bits in objects to be able to implement
[lightwieght locks](https://webkit.org/blog/6161/locking-in-webkit/).

Yet another complication, is that on 32bit systems, the `shape_id` is only 16bits long.
I have the project to use the upper bits of the `shape_id` to store metadata, such as the `frozen` and `too_complex`
status, allowing to test for this without chasing a pointer: https://github.com/ruby/ruby/pull/13289.
But this currently can't be done on 32bit sytems, both because accessing the `shape_id` might require a hash-lookup,
and also because it's only 16bits long, so every single bit used for tagging severely restrict the maximum number of
shapes.

### Proposal

To simplify all this, we propose that on 32bit systems, we add a `VALUE shape_id` in `RBasic`:

```c
struct RBasic {
    VALUE flags;
    const VALUE klass;
#if RBASIC_SHAPE_ID_FIELD
    VALUE shape_id;
#endif
}
```

This ensure that on 32bits, all objects have their `shape_id` always at the same predictable offset, and 32bits long.

As you can see on the pull request, it simplify the code quite significantly: https://github.com/ruby/ruby/pull/13341,
and there's more cleanup that can be done.

The downside obviously is that on 32bit, objects would grow from `20B` to `24B`.

### Pull Request

You can find the proposed patch at: https://github.com/ruby/ruby/pull/13341

cc @tenderlovemaking and @jhawthorn 

Also FYI @katei because as the maintainer of WASM I assume this impact you the most.




-- 
https://bugs.ruby-lang.org/
 ______________________________________________
 ruby-core mailing list -- ruby-core@ml.ruby-lang.org
 To unsubscribe send an email to ruby-core-leave@ml.ruby-lang.org
 ruby-core info -- https://ml.ruby-lang.org/mailman3/lists/ruby-core.ml.ruby-lang.org/