Factor/Dynamic sealing

This page is about sealing with dynamic extent. One of the primary intended uses is for Metacircularity, but any situation where the programmer wants to close the polymorphic semantics of code in order to achieve very high, stable performance would benefit from this approach.

Concepts

Sealing

This is a specific kind of restriction to the language. The name is borrowed from Dylan. While my original concept was first designed without knowing the Dylan approach (that I don't know in deep details) I'm now trying to gather any wisdom that might come Dylan. If you think you can help, please contact me on the Concatenative IRC channel (prunedtree).

In essence, sealing means to close some subsystem from further changes. This allows the compiler to do full code analysis and remove any overhead from polymorphism in the limit. As sealed code is meant to have high, stable performamce, it is expected that the compiler will very aggressively optimize it, taking lots of time and producing lots of code if needed. That cost is bounded for several reasons: Because we seal a core subset of the language that is highly optimizable, and because sealed code has, by definition, far less dependencies, it seldom needs to be recompiled.

What is important to understand is first that sealing is a very light restriction: We close the system, but we keep all high level features of the language (that make it so much better than C, even for high level code: managed memory, quotations and polymorphism, among others). Also, sealing offers us an important guarantee for complex, low level code: Only code within the sealed core system will have an effect on our code, which prevents spurious bugs (or worse, severe random performance regressions) to appear because of random user-level code.

So what do we loose ? If only the sealed core system is running... nothing. What we loose is the possibility to extend the sealed code with arbitrary user code (which is what we desire in this case). Note that all the sealed core system can be used with arbitrary extensions when called from user code. This is desirable, and this is why we use dynamic sealing, not static sealing

Dynamic

Dynamic has the meaning of dynamic scope here, not 'dynamic language'. Dynamic scope simply means the scope is in sync with the runtime stack, as if the dynamic state was passed as an argument to every call. The dynamic state here is the 'access control' object. That object is used for subjective dispatch in order to enforce sealing.

Subjective Dispatch

Dynamic sealing is implemented as a very stable (in the sense that the subject doesn't change often) form of subjective dispatch. Subjective dispatch is a simple extension to multiple dispatch were we dispatch on an additional dynamic variable, the subject. because the subject has dynamic extent, you can consider that this is equivalent to adding a 'subject' argument to all functions in the system, and dispatching on it's class for all generic functions.

Usage and 'why it works'

This revision created on Wed, 22 Oct 2008 07:44:54 by prunedtree