Type Specialized Instance VariablesΒΆ

Posted: July 14, 2013

In Topaz, like most other VMs, all objects are stored in what are called “boxes”. Essentially that means when you have something like x = 5, x is really a pointer to an object which contains 5, not the value 5 itself. This is often a source of performance problems for VMs, because it generates more garbage for the GC to process and means that to access the value 5 more memory dereferences are needed. Topaz’s just-in-time compiler (JIT) is often able to remove these allocations and memory dereferences in individual loops or functions, however it’s not able to remove them in structures that stick around in memory, like objects.

Therefore, over the past week I’ve been working on an optimization for Topaz called “type specialized instance variables”. Basically what that means is that Topaz keeps track of what types instance variables in an object tend to have, and then specializes the storage to remove the indirection for Fixnum and Float objects.

Let’s look at an example:

class Point
  def initialize(x, y, z)
    @x = x
    @y = y
    @z = z

p = Point.new(1, 2, 3)

Before this optimization, p looked like this in memory. Each box indicates an 8-byte (on 64-bit systems) value, and arrows are pointers:


And after the optimization, it looks like this:


What are these maps? They’re a concept out of SELF-88, basically they describe the layout of the object, plus some JIT magic so it’s free to find the position and type of a field. Before this patch the map looked like:

    "x": {"position": 0},
    "y": {"position": 1},
    "z": {"position": 2}

After this patch it looks like:

    "x": {"position": 0, "kind": "int"},
    "y": {"position": 1, "kind": "int"},
    "z": {"position": 2, "kind": "int"},

This is dramatically simplified, if you’re interested in the full details of how they work, it’s very similar to how they work in PyPy.

With this optimization landed, Topaz will use less memory and be faster for programs that store Fixnum and Float objects in memory. If you’re interested in this type of optimization you can read about a similar one in PyPy for lists that we’re in the process of porting to Topaz.

We’re looking forward to doing our first release soon, we hope you’ll test Topaz out, and give us feedback with the nightly builds until then

Previous topic

One Year of Topaz