Feature #16984

Remove write barrier examption for T_ICLASS

Added by alanwu (Alan Wu) about 1 month ago. Updated 24 days ago.

Target version:


Consider the following code:

module M
  def foo; end
  def bar; end

class C
  include M

The object reference graph from running the code looks like this:

+---+              +-----+
| M |--------------| foo |-+
+---+              +-----+ |
  |                +-----+ |
  +----------------| bar | |
                   +-----+ |
+-----------+         |    |
| iclass(M) |---------+    |

Applying the proposed patch, the graph becomes

+---+         +--------------+   +-----+
| M |---------| method table |---| foo |
+---+         +--------------+   +-----+
+-----------+         |    |     +-----+
| iclass(M) |---------+    +-----| bar |
+-----------+                    +-----+

This change has a similar effect on the constant table. In addition to this, T_ICLASS no longer
holds a reference to a ivar table. Code that access the ivar table through iclasses
are changed to access it through the object from which the iclass was made. This change
impacts autoload and class variable lookup.


The main goal of this change is to make iclasses and modules write barrier protected. At the moment, they are
"shady", which means the GC has to do extra work to handle them. In code bases that use modules a lot,
iclasses can easily take up a significant portion of the heap and impact GC time. Inserting write barriers was
tricky in the old setup, because of the way M and iclass(M) share the method table.

Having write barrier for iclasses mean they can age in the generational GC.
Once aged, the GC can sometimes skip subgraphs rooted at these objects, improving performance.

Impact to GC time

I measured the impact to minor GC time with the following steps:

  • load an application
  • run GC::Profile.enable
  • allocate 50 million objects
  • run

Here is the impact to average minor GC time on various apps:

Application Before After Speedup ratio
CRuby's test-all suite 2.438ms 2.289ms 1.06
rails new app 1.911ms 1.798ms 1.06
Private app A 5.182ms 5.168ms 1.00
Private app B 185.7ms 107.9ms 1.72

Private app A's heap size is about 22 MiB compared to B's 250 MiB.
App B boots up about 15% faster with this change.

Impact to class variable lookup

I included a benchmark in the patch to measure the impact to class variable lookup performance.
The difference seems negligible.


This change seems to reduce minor GC time for real-world applications.

Credits to tenderlovemaking (Aaron Patterson) for coming up with the idea for this change.

Updated by ko1 (Koichi Sasada) about 1 month ago

Thank you for great work. This kind of hack can cause BUGs easily.

Private app A's heap size is about 22 MiB compared to B's 250 MiB.

Could you measure the memory/objects consumption before and after this patch if it is not difficult?
Maybe no problem, but I want to confirm.

Updated by alanwu (Alan Wu) about 1 month ago

Could you measure the memory/objects consumption before and after this patch if it is not difficult?

I took measurements on app B. It's a large Rails app with lots of classes and modules.
The amount of retained memory is not deterministic unfortunately, so I can only give a rough summary.

Change to median Change to average
GC::Profiler "Total Size" 7 MiB 1 MiB
VmRSS from the /proc 4 MiB -5 MiB

The increase to GC heap size makes sense, I expect about 5 MiB more objects given the number of
classes and modules in the app. There is not much change to RSS I guess because the
patch moves what used to be on the malloc heap to the GC heap.

Updated by alanwu (Alan Wu) 24 days ago

  • Description updated (diff)

Edit: I noticed that T_ICLASS wasn't marking the shared method and constant table on master. My notes
about reducing the number of gc_mark references on the heap was incorrect.

Also available in: Atom PDF