When methods with extended specializers are compiled, a style-warning like the following is signalled:

style-warning:
  can't find type for specializer (ACCEPT "image/webp") in
  SB-PCL::PARAMETER-SPECIALIZER-DECLARATION-IN-DEFMETHOD.

Fixing this probably requires further changes in PCL.

I think it would be really nice if class and quote pattern specializers (without any further restrictions or variable bindings) were treated identically to the standard class and eql-specializers. This would allow the brave user to default to using pattern generic functions, and only paying for the cost when a non-trivial pattern specializer is actually included on a generic function.

As I wrote earlier, pattern-more-specific-p establishes a partial order
on patterns which is similar to the subtype relation (See attached
figures for examples; arrows point to "more specific"). The similarity
being that the set of values matching a more specific pattern is a
subset of the values matching a less specific pattern.

Furthermore, the dispatch strategy (considering only one argument and
only pattern specializers) currently looks like this (unchanged compared
to previous email):

1 for each "specializer cluster" C (explanation below), in any order:
2       for each pattern P in C, most specific first:
3               if P matches the argument:
4                       return generalizer object with
5                               all specializers in the "cluster" whose 
6                               pattern is P or a less specific pattern

The rationale being that a single successful match should be sufficient
to determine all matching patterns (and thus accepting
pattern-specializers) as well as binding the relevant pattern variables.
This assumes that optima is smart enough to avoid unnecessary work when
sequentially trying subsequently less specific patterns (in line 2).

Now the new issue: previously, I assumed patterns would form totally
ordered "clusters" (actually connected components) under
pattern-more-specific-p. This assumption is reflected in the algorithm
above. However, the assumption is not true (again, see attached figures,
ignore BUILT-IN-CLASS CONS …).

A new assumption could be that patterns form connected components in a
DAG under pattern-more-specific-p. If I'm not mistaken, a suitable
adaptation of the above algorithm would then be:

1 for each connected component C, in any order:
2       for each pattern P in C, in topological order, most specific first:
3               if P matches:
4                       return generalizer object with
5                               all specializers whose pattern is in the 
6                               transitive PATTERN-LESS-SPECIFIC-P-closure
7                               of { P }

Not sorting connected components for processing should have performance
implications but should not affect the result of the computation since
the sets of matching values should be disjoint.

One potential problem is that pattern-more-specific-p employs subtypep
when comparing (type …) and class patterns (see specializer-dag-2 for
excessive use of class patterns), making the resulting ordering

1. Independent of generalizer objects (convenient, maybe good)
2. Different from CLOS semantics (maybe bad)

specializer-dag-1:

specializer-dag-2:

I would like to resolve the tiebraker TODO as follows:

• If the same media type entry is the best match for two specializers (e.g. request only has text/* and there are specializers text/plain and text/html), use string< on the specializers' media type strings.
• If two media types in the accept specification, that do not have a ancestor-descendent relation in the accept-tree, have the same q value, treat the media type that occurs earlier in the specification as preferred.

I have implemented this and could push if the approach sounds reasonable.

We have two tasks at generic function invocation time:

1. Turn the list of required arguments into a list of generalizers
2. Compute a list of "effective arguments" for the effective method (e.g. pattern specializers inject variable bindings at this point)

The first task is handled via either

• generalizer-of-using-class gf arg

  Has the advantage of simplicity

• or a combination of generalizers-of-using-class gf args and generalizer-of-using-class gf arg

  Has the disadvantage of complexity but is more powerful and can be more efficient. For example, the pattern-specializer implementation has a method which basically does (mapcar #'funcall list-of-generalizer-computation-functions args) without generic function calls for individual arguments.

Potential improvements:

• For the first approach, change generalizer-of-using-class gf arg to generalizer-of-using-class gf arg arg-position to allow processing based on the position of the argument. This would allow performing processing of arguments based on their position like in the pattern-specializer example above.
• Change generalizers-of-using-class gf args to generalizers-of-using-class gf args num-required to avoid consing up the list of required arguments at invocation time. This can be applied to both approaches.
• The current implementation can be improved by closing over the number of required arguments in the discriminating function instead of computing it at invocation time.

The second task is not currently handled in the master branch

One solution could be a new protocol function compute-effective-arguments-function gf num-required which would return a function with lambda-list (args generalizers) which would in turn produce the "effective argument list" when called, for example by injecting additional arguments.