Thu Dec 31 15:10:41 CET 2009

VM primitives


- value store with explicit boxes (set! and mutable varrefs are transformed)

- all global variables are boxed.  code can be optionally "frozen"
  when compiled, to give a poor man's module system.

- environment is a list of vectors.  use GC integers to index frame+var

- continuations frames (data structure) are tagged with functions that
  take a single input argument and update the machine state, and with
  a dictionary of continuation marks.


 * I wonder if the PF VM and the Schem VM can be unified.  I guess..
   If the current environment is equated to the data stack?  Yes let's
   do that.

 * A value store allows (partial) elimination of the environment at
   run time: variables that have known bindings can be substituted.

 * How to properly distinguish un-evaluated code from values = result
   of evaluation?

 * What is a continuation frame for an application?  Can this directly
   construct the environment?  I.e. always first evaluate the function
   closure, and then simply concatenate.

 * The question is really: what is a closure, thinking of PF?  

   It's something that takes arguments, and works like this:
     - replace the current data stack with the one embedded in the closure
     - evaluate arguments and push them on the data stack
     - jump to the closure's code

   I.e. it's a PF function that _replaces_ the data stack before it is

   Translating the above to a machine instruction sequence:
     - make a continuation frame that restores the current environment

Let's try to build this.  This seems to be an interplay between
instructions and continuations, which can both be represented as
primitive (C) functions.

Code is an expression tree (no sequential VM), represented by cons

(prim . args)

All highlevel function closures are defunctionalized, and passed to an
apply primitive.

The machine primitives are CPS: they do not return, but instead update
the machine state.

Looks like plain and simple CPS is preferrable (see EOPL).  The
non-returning lambda expressions can then be represented as structs.

So how are CPS and DF related?

From Danvy's paper[1]: DF can be used to represent the closures
resulting after performing CPS transform.

The main intuitive stumbling block is 'if'.  Somehow I don't understand
why this needs two primitives: one to start the evaluation of the
condition, and one to pass control to the proper expression based on
the boolean value of the condition.

PLAN: It looks like what I'm about to do is quite close to chapters
8,9,10 in EOPL, so it might be a good idea to simply read those again.

[1] http://www.brics.dk/RS/01/23/