Translated from matz blog, 2005-02-21:

This report is old, but I didn’t know.

For implementing Dynamic Dispatch, the technique that uses the dispatch table like Virtual Table(VTABLE) of C++ is often used in statically-typed language. This is because a statically-typed language has properties such as “fast” and “abundant infomation on compile”.

The above-mentihoned report are written about applying such a dispatch table technique to a dynamically-typed language like Smalltalk. It’s considered about the point how it deals when a class or method is dynamically added.

Novel points are:

• It compress two dimensional tables(class x selector) to one dimensional one. It sorts a sparse table, and it stores elements in desc order. It uses the offset to get.
• When compressing the table, though it is usually splitted by class, this report recommends to split by selector. Therefore, the efficiency when updating is up because adding a selector ends only adding a line.
• When a class is added, it deals by using two or more dispatch tables (fall-back the 2nd table by method_missing). Though a new class is a little slow, it’s possible to gain a whole perfomance. In an idle moment, it restructures dispatch tables.

The effeciency seems to be quite good. However, it may be that this technique is inefficient on like Ruby that has a open class, an existing class can be added the method, by increasing table-restructuring.

I think how to solve it.

Hopefully things will only be on hold for a day longer. I’m trying to work out my problems with DRb on FreeBSD. It’s been a pain to restart the file-sharing service every couple hours. With some luck, we’ll continue from Tuesday through Friday.