source: branches/streams/abcl/doc/design/amop/d-m-c-notes

Below is a DRAFT e-mail that I intend to send to the mailing list,
however, having it in the repository (probably rephrased) is better
long-term documentation.



Over the past days, I've been working on porting SBCL's D-M-C tests to ABCL's
test suite and testing+fixing our implementation. A number of use-cases have
been fixed, however, I'm now down to the more complex cases, in particular
the case for the (:arguments . lambda-list).


Context
-----------

When handling EMF computation, there are two sets of arguments (lambda lists):

1. the arguments passed to the METHOD-COMBINATION through the
   (:method-combination ...) form in the generic function definition
2. the arguments passed to the generic function when it is being called

This distinction is very important, yet not particularly clear from our
sources. The former set of arguments is available from the instantiation of
the generic function (DEFGENERIC evaluation) and constant throughout the life
of the GF. The latter is set of arguments is not available until the function
is being called and will presumably be different for each invocation of the GF.

The former set is passed to the D-M-C form in the second position:
 (D-M-C <name> <arguments> ....). The latter set is made accessible by
providing the (:arguments ...) form to the D-M-C form -- binding of the
variables happens at "EMF-calculation-time".

Current implementation
---------------------------------

Our existing implementation does not work at all with the (:arguments ...)
option in the D-M-C definition. [SBCL didn't either, btw,
until 0.7.<something>]
What happens in our implementation is that the function
STD-COMPUTE-EFFECTIVE-METHOD-FUNCTION calls a function created by the D-M-C.
That function returns forms to be used as the EMF. S-C-E-M-F wraps the returned
forms in a function and returns it as the EMF.

This works as long as the EMF does not depend on the arguments supplied
to the GF (generic function) call.


The problem
------------------

Our implementation tries to access the function call parameters (resulting
in "unbound variable errors") from the EMF-generating function. However,
that function won't (ever) be passed the call arguments.


The solution
-----------------

Writing down the above and taking into account that we want to cache as much
of our EMF as possible for performance reasons as well as considering that
the EMF depending on the function call arguments can't be cached, I think
this is the solution:

The forms being returned (and later wrapped in a lambda) should include code
which does another code-generation step --with access to the call parameters--
and include a call to the forms having been generated.

Examples
--------------

A call to the EMF-generating function which does not depend on the call
arguments would return something like:

 '(CALL-METHOD (MAKE-METHOD (error "ABC 123")))

This form will be wrapped in a lambda roughly like this:
  (lambda (args) (macrolet ((call-method ...)) <forms>))


A call to the EMF-generating function which *does* depend on the arguments
would return something like: