Re: After GOOPS integration: Computation with native types!

[Mikael Djurfeldt <mdj at mdj dot nada dot kth dot se>]

Keisuke Nishida <kxn30@po.cwru.edu> writes:

> Do I need to include some special feature in my VM?  Hmm, but maybe
> I shouldn't do that now...

Probably not, so I probably shouldn't answer, but...  :)

You'll need to include some extremely efficient mechanism to do
multi-method dispatch.  The SCM_IM_DISPATCH form, with its
implementation at line 2250 in eval.c, is the current basis for
efficient dispatch in GOOPS.

I think we should develop a new instruction for the VM which
corresponds to the SCM_IM_DISPATCH form.

This form serves both the purpose to map argument types to the correct
code, and as a cache of compiled methods.

Notice that I talk about cmethods below, not methods.  In GOOPS, the
GF has a set of methods, but each method has a "code-table" mapping
argument types to code compiled for those particular concrete types.
(So, in essence, GOOPS methods abstractly do a deeper level of type
dispatch.)

The SCM_IM_DISPATCH form has two shapes, depending on whether we use
sequential search (few cmethods) or hashed lookup (many cmethods).

Shape 1:

 (#@dispatch args N-SPECIALIZED #((TYPE1 ... ENV FORMALS FORM1 ...) ...) GF)

Shape 2:

 (#@dispatch args N-SPECIALIZED HASHSET MASK
             #((TYPE1 ... ENV FORMALS FORM1 ...) ...)
             GF)

`args' is (I hope!) a now historic obscure optimization.

N-SPECIALIZED is the maximum number of arguments t do type checking
on.  This is used early termination of argument checking where the
already checked arguments are enough to pick out the cmethod.

The vector is the cache proper.

During sequential search the argument types are simply checked against
each entry.

The method for hashed dispatch is described in:

http://www.parc.xerox.com/csl/groups/sda/publications/papers/Kiczales-Andreas-PCL

In this method, each class has a hash code.  Dispatch means summing
the hash codes for all arguments (up til N-SPECIALIZED) and using the
sum to pick a location in the cache.  The cache is sequentially
searched for an argument type match from that point.

Kiczales introduced a clever method to maximize the probability of a
direct cache hit.  We actually have 8 separate sets of hash codes for
all types.  The hash set to use is selected specifically per GF and is
optimized to give fastest average hit.


What we could try to do as soon as the VM is complete enough is to
represent the cmethods as chunks of byte code.  In the current GOOPS
code, the compilation step (which is currently empty) is situated in
`compile-cmethod' in guile-oops/compile.scm.  [Apologies for the
terrible code.  That particular part was written at Arlanda airport
after a sleepless night (packing luggage, not coding), on my way to
visit Marius (who, BTW, didn't take GOOPS seriously.  ;-)]