Feature #15765

Updated by alanwu (Alan Wu) 4 months ago


The current implementation of `Module#name` is known for having sub-optimal performance when it comes to anonymous modules. (see #11119 and #15625)
I have put together a reimplementation of `Module#name`, which works by eagerly naming
modules and classes when they are assigned to constants. Besides solving the performance
issues for `Module#name` on anonymous modules, there are some other benefits.

This patch:
- removes more code than it adds
- makes normal class and module definition slightly faster (definitions like `class Foo; end`)
- slightly reduces memory usage for classes and modules due to the removal of a hidden ivar
- improves the performance of defining modules and classes under an anonymous module. This used to execute a global search each time. search.

### Behavior changes and caveats:

Since we already name module and classes declared with the `class` and `module` keyword on trunk, this patch mostly targets anonymous
modules. I tried my best keeping the behaviors consistent with the current implementation, but there are some small behavioral changes.

mod =
mod::BeforeToS =

mod.to_s # on trunk, the VM starts naming modules assigned under mod after calling to_s

mod::AfterToS =

p # nil on both
p # nil on both

p # "#<Module:0x0000563494b1cca8>::AfterToS" on trunk, nil after patch
p # "#<Module:0x0000563494b1cca8>::AfterToS2" on trunk, nil after patch
This prints 4 nils after my patch, as I think the behavior on trunk is unintentional. A few C APIs also have the same effect as calling to_s. They are all changed to be side-effect free.

m =
m::Child =
Mod = m
p Object.send(:remove_const, :Mod)
This prints nil on trunk and `Mod::Child` under this patch.

`rb_name_class` is removed, as it does nothing in this new implementation. Not sure if this is public API.

Since the recursive naming is done with a recursive function, when function. When a deeply nested anonymous module is assigned
to a constant, it is technically possible for this implementation to throw a `StackError`. I had a version
which does heap allocation to deal with this, but I picked this version for performance in the common cases.
Anonymous modules are rare as is, and one would have to build a structure nested thousands level deep for this to happen.
On my system it can name a module fifty thousand levels deep without problem.

I think these changes are fairly minimal and acceptable.