Notes
================================================================================
Subsumption
--------------------------------------------------------------------------------
If p1 subsumes p2 it needs to be sorted so that it comes later.
First Example
--------------------------------------------------------------------------------
(match [x y z]
[_ false true ] 1
[false true _ ] 2
[_ _ false] 3
[_ _ _ ] 4)
Heuristics
--------------------------------------------------------------------------------
First Row
f(i) = 0 if _
f(i) = 1 otherwise
Small default
-v(i) where i equal number of wildcard patterns in column i
b(i) branching factor - negative number of switch nodes
a(i) negation of the sum of the arities
l(i) number of children of the emitted switch
r(i)
q constructor prefix
qba
q 1) avoid usefulness computations, 2) avoid pattern copies
p(j i) is a generalized constructor pattern test
Hmm sounds like we don't need to do usefulness checks with qba?
Useless clauses
--------------------------------------------------------------------------------
Subsumption checks to delete rows, before the subsumption check as well.
If we can throw away redundant checks and move forward
good. Non-exhaustiveness is not really something we care about that
much - I don't think. People either added the default handler or their
didn't.
Series of examples are excellent at the end of "Warnings for pattern
matching", should definitely work against these.
Reducing necessity to usefulness.
type mylist = Nil | One of int | Cons of int * mylist
New Thoughts
--------------------------------------------------------------------------------
| x (isa? x A) _ _ |
| _ y (isa? y B) _ |
| x (isa? x C) y (isa? y C) 0 | x :guards [(isa? x D)]
*{0}
If there are guards remaining on x that's another pattern matrix, other people
might use that guard as well, if it's an isa? or = check we can save space.
During compilation we could just use an atomic integer to associate an id with
a particular class when *compile-files*
Thoughts
--------------------------------------------------------------------------------
We're going to have to spend a couple of days just sorting out an
optimal data representation.
Design
--------------------------------------------------------------------------------
Interesting case:
[x y] :guard [(odd? x) (even? y)] :one
[x 0] :two
[0 y] :guard [(odd? y)] :three
[_ _] :four
As matches are added we'll need to recompute the optimal matrix. We'll need to
consider how to make this fast, as well what we can do minimize the
work.
Step1
--------------------
[x y] :guard [(odd? x) (even? y)] :one
{:guards {[0] [(odd? x)]
[1] [(even? y)]}
:patterns [{:arglist [x y] :guard [(odd? x) (even? y)]}]
:start 1
:dag {0 ::no-method
1 {:test (odd? x)
:edges {true 2 false 0}}
2 {:test (odd? y)
:edges {true 3 fales 0}}
3 :one}}
Step2
--------------------
[x 0] :two
{:guards {[0] [(odd? x)]
[1] [(even? y)]}
:patterns [{:arglist [x y] :guard [(odd? x) (even? y)]}
{:arglist [x 0]}]
:start 1
:dag {0 ::no-method
1 {:test (odd? x)
:edges {true 4 false 0}}
2 {:test (odd? y)
:edges {true 3 false 0}}
3 :one
4 {:test y
:edges {0 5 ::default 2}}
5 :two}}
We need something like insert-node, which will shift a node.
Step3
--------------------
[0 y] :guard [(odd? y)] :three
{:guards {[0] [(odd? x)]
[1] [(even? y)]
[1] [(odd? y)]}
:patterns [{:arglist [x y] :guard [(odd? x) (even? y)]}
{:arglist [x 0]}
{:arglist [0 y] :guard [(odd? y)]}]
:start 6
:dag {0 ::no-method
1 {:test (odd? x)
:edges {true 4 false 0}}
2 {:test (odd? y)
:edges {true 3 false 0}}
3 :one
4 {:test y
:edges {0 5 ::default 2}}
5 :two
6 {:test x
:edges {0 7 ::default 1}}
7 :three}}
We'll have to rewrite parts of the graph, but that means we'll
have to the parts of the graph that return.
Step 4
--------------------
[_ _] :four
{:guards {[0] [(odd? x)]
[1] [(even? y)]
[1] [(odd? y)]}
:patterns [{:arglist [x y] :guard [(odd? x) (even? y)]}
{:arglist [x 0]}
{:arglist [0 y] :guard [(odd? y)]}
{:arglist [_ _]}]
:start 6
:dag {0 :four
1 {:test (odd? x)
:edges {true 4 false 0}}
2 {:test (odd? y)
:edges {true 3 false 0}}
3 :one
4 {:test y
:edges {0 5 ::default 2}}
5 :two
6 {:test x
:edges {0 7 ::default 1}}
7 :three}}
We still haven't considered the problem of necessity. This really feels
like we something *before* the dag, that we can use to generate the
dag.
| (odd? x) (even? y) |
| x 0 |
| 0 (odd? y) |
| _ _ |
Aha. The column representing is actually better!
| 0 (odd? y) |
| (odd? x) (even? y) |
| x 0 |
| _ _ |
We should put the example from Efficient Predicate Dispatch into this
format.
When there is an internal access, should just be considered as another column?
| 0 _ (odd? y) |
| (odd? x) _ (even? y) |
| x _ 0 |
| (A. y) y _ |
| _ _ _ |
User defines the following method
(defm foo [x] :guard [(even? x)]
:two)
We now have a new method in the method-table called myns/foo, this
method entry will look like the following:
{:guards {[0] [(even? x)]}
:patterns [{:arglist [x] :guard [(even? x)]}]
:start 0
:dag {0 ::no-method
1 {:test even?
:edges {true 1 false -1}}
2 :two}}
User then defines the following addition to the method
(defm foo [0] :one)
This will update the method entry in the table it should now look like
the following:
{:guards {[0] [(even? x)]}
:patterns [{:arglist [x] :guard [(even? x)]}
{:arglist [0]}]
:start 2
:dag {-1 ::no-method
0 {:test (even? x)
:edges {true 1 false -1}}
1 :two
2 {:test x
:edges {0 3 false 0}}
3 :one}}
This works for the simple case, but what if the reordering happens in
the middle?
(defm foo [4]
:three)
This isn't a problem as this is just another equality test:
{:guards {[0] [(even? x)]}
:patterns [[x] [0] [4]]
:start 2
:dag {-1 ::no-method
0 {:test (even? x)
:edges {true 1 false -1}}
1 :two
2 {:test x
:edges {0 3 4 4 false 0}}
3 :one
4 :three}}
So this is pretty easy to deal with.
Can we think of a situation that would actually cause a reordering
that we would want to handle? Can't really think of any cases that
involve Clojure.
[(A. 0)]
[(B. 0)]
[(C. 0)]
A -> B -> C hierarchy
Concrete type matches are not a problem, they're always more
specific.
[x] :guard (isa x IFoo)
[x] :guard (isa x IBar)
x is A which implement IFoo and IBar
If we know the type of x we can point out the ambiguity.
At the bottom the user could add, after :where,
:prefers [[IBar IFoo]]
I don't suppose this'll be any worse then
Ideas
--------------------------------------------------------------------------------
In the cases where one branch has many more tests, I guess a cascade
of cases still wins out really over if
(supers (class 0))
#{java.lang.Number java.io.Serializable java.lang.Object
java.lang.Comparable}
This is incredibly useful information for dispatching.
Issues
--------------------------------------------------------------------------------
Getting a specific method. Not sure how important this is, we would want to
preserve the destructuring perhaps? Or maybe you get a function that
is closed over the arguments that you gave to it? A method signature
is its pattern and it's guard
Algorithm
--------------------------------------------------------------------------------
patterns are two things, a type and a series of bindings. Unlike
destructuring in Clojure, we want to differentiate between things we
want to treat as sequences and things that are of fixed length.
() ; always represents fixed length sequences, ISeq + count
[] ; the usual clojure interpretation, ISeq
{} ; the usual clojure interpretation, IMap
(A. f1 f2 ...) ; field syntax
We test the arguments in order, we grab all the tests for the first
argument organized by the method they dispatch on. The pattern plus
the normalized guard is the unique key (we should probably check for
alpha equivalence) for each method.
We could also use the strategy outlined by the OCaml paper, using the
concept of the necessity. The math is a bit thick in that paper, would
need some help?
Pattern Matrix
--------------------------------------------------------------------------------
In the map there is a (n x m) matrix - but our tests are not just (n x
m), some patterns may require more matches since we have guards. Need
to read the paper several more times to suss out it's meaning, the
mathematics is not very complicated. It's mostly a notation to
simplify the description w/o getting into the details of OCaml.
P1.1 P1.2
P2.1 P2.2
P3.1 P3.2
Perhaps best if all tests are considered as one particular
pattern. Different kinds of tests are given a different priority.
P.1.1 = (and (type Foo a) (= (:x a) 0))
P.2.1 = (and (type Bar a) (= (:y a) 1))
P.3.1 = (and (type Foo a) (even? (:x a) 0))
There's recursive stuff going on here. P1.1 and P3.1 kind of represent
a sub-decision tree.
Heuristinc & Scoring
--------------------------------------------------------------------------------
f = score constructors 1
v(i) = score wilcard occurrences as -v(i)
Integration with a Type System
--------------------------------------------------------------------------------
Hmm, it would be interesting to use the logic engine to store a
database of facts as your declaring you're defining your program and
emit type warnings. We would even have to interfere with the compiler,
we'd simply run a clojure parser on the body and run the type
inference. The fact that definitions have to be defined in a specific
order really simplifies the type checking problem.
Racket
--------------------------------------------------------------------------------
Racket supports recursive pattern, this is interesting. Steps towards
Datalog yet again?
A big benefit of the inability to have circular references in Clojure
is completely avoid the infinite loop issue.
Classes/Interfaces/Protocols - many possible branches
boolean tests - 2 branches
[x y]
A (> x 1)
B (= x :foo)
C A
What other things would people want to dispatch on that are
legitimately multiway?
structural matching, whatever the defrecord literal becomes when it comes.
(A. 1 2)
[{:keys [a b] :as b}] :guard ....
The problem here that we have a two steps.
for things which we know are types, we can emit case
for things which we know are ad-hoc, we can emit condp isa?
Use cases
--------------------------------------------------------------------------------
We would like to know that the following expressions are equivalent.
(= (:foo x) 'a)
(= (:foo x) 'b)
(= 'c (:foo x))
(= 'd (:foo x))
If we know they are then we can setup a multiway branch again.
This simple case can definitely be handled.
We could support Racket like things.
(? every? even? foo)
Predicate Dispatch
--------------------------------------------------------------------------------
[{f1x :x f1y :y :as f1} {f2x :x f2y :y :as f2}] :guard [(isa? f1 A)
(isa? f1x A)
(not (isa? B f1))
(isa? f2x C)
(= f1y f2y)]
[{f1x :x :as f1} f2] :guard [(isa? f1 B) (isa? f1x B)]
[{f1x :x :as f1} f2] :guard [(isa? f1 C) (isa? f2 A) (isa? f1x B)]
[f1 f2] :guard [(isa? f1 C) (isa? f2 C)]
[f1 _] :guard [(isa? f1 C)]
| f1#A f2 |
| f1#B f2 |
| f1#C f2#A |
| f1#C f2#C |
| f1#C f2 |
NOTE: how to deal with multiple isa? on a subterm? (isa? x B) (isa? x A)
column 1 is clear the most important since it tests for type in all cases
This is actually much simpler than what is suggested in the Efficient
Predicate Dispatch paper, if f1#A is selected we don't need to use
those tests anywhere else, we can can remove predicates as we
specialize and if there are no other matches, we can just destructure
and run the remaining bits?
Same for B
For C we actually have some legitimate tests
Issues
-----------------------------------------------------------------------------
One thing that is tough is recovering the method.
(get-method foo [0 1])
Should return the method that matches the particular argument
list. Need to think about this.