Constant applicative form
Also referred to as CAF.
Any super combinator which is not a lambda abstraction. This includes truly constant expressions such as
12, ((+) 1 2), [1,2,3] as well as partially applied functions such as
((+) 4). Note that this last example is equivalent under eta abstraction to
\ x -> (+) 4 x which is not a CAF.
Examples and discussions
Since a CAF is a super combinator, it contains no free variables. Moreover, since it is not a lambda abstraction it contains no variables at all. It may however contain identifiers which refer to other CAFs, e.g.
c 3 where c = (*) 2
A CAF can always be lifted to the top level of the program. It can either be compiled to a piece of graph which will be shared by all uses or to some shared code which will overwrite itself with some graph the first time it is evaluated. A CAF such as
ints = from 1 where from n = n : from (n+1)
can grow without bound but may only be accessible from within the code of one or more functions. In order for the garbage collector to be able to reclaim such structures, we associate with each function a list of the CAFs to which it refers. When garbage collecting a reference to the function we collect the CAFs on its list.
Q&A discussion from HaWiki
- What is the difference between
\x -> (+) 4 x, apart from the obvious difference in notation? (I have braced the + to make the expressions compliant with Haskell syntax.)
- One is an application that returns a function and may imply some work in general, the other is a syntactic value (a lambda abstraction).
- If notation really happens to matter so much, then what about
(+ 4)? Are both of these CAFs? If we take away the nifty syntax, how do we express
(+ 4)then? Or is it no longer a CAF? ;)
- It depends on how they desugar. `(4 +)` presumably desugars to `((+) 4)` and so is a CAF. If (+ 4) desugars to `\x -> x + 4` it isn't, if it desugars to `flip (+) 4` it is. -- DerekElkins
- --- Rambler I'm afraid I still don't get it (or maybe I do, but I secretly wish it didn't mean that...) To the best of my understanding of the matter, `flip (+) 4` should return something that should behave exactly as `\x -> x + 4` would. Moreover, even if I take it for granted that `flip (+) 4` differs dramatically from `\x -> x + 4`, what happens when we expand the former?
- There is no difference in behavior as far as meaning is concerned. They are just treated differently. -- DerekElkins
- Suppose `flip` is defined as `\f x y -> f y x` (which it probably is, anyway,) then `flip (+) 4 where flip = \f x y -> f y x` is obviously not a CAF, according to the definition.
- What about `flip (+) 4 where flip f x y = f y x`???
- I believe that FOLDOC is incorrect here. A better way of thinking about CAFs is that a CAF is anything that can be let floated (see lifting pattern) to the top level. So a CAF can not only call other CAFs, but also other super combinators.
- The difference between
flip (+) 4and
\x -> x + 4is that in a pure lambda calculus implementation, the former can be reduced but the latter cannot. Remember that
\f x y -> f y xis a shorthand for
\f -> (\x -> (\y -> f y x))), which means if given two arguments, it can be reduced. --AndrewBromage
- Fine, is it then irreducibility that makes CAFs special? If so, then the current formulation of the definition clearly misses the point. If not, then why is this distinction relevant in the context of the definition?
- In any case, I am left wondering 1) why irreducible expressions might be of such interest (apart from the trivially imaginable,) and 2) whatever happened to the proverbial referential transparency. It just struck me lately that my difficulty may be due to the fact that I am trying to make sense of CAFs in a broad theoretical context, whereas they may be simply an implementation instrument -- is this the case? --Rambler
- You've inverted it slightly (or it reads that way): it's reducibility that makes CAFs special. Since they are reducible they may imply work; work that could be saved. Nevertheless, yes, one typically doesn't care whether something's a CAF or not as far as meaning goes. Read the relevant parts of the book I linked to below, if you haven't already. As for ReferentialTransparency, whether or not something is viewed as a CAF doesn't change it's meaning, i.e. it is just a different implementation of the same function. -- DerekElkins
- OK, it definitely makes sense now, thanks to everyone for the explanations and references. --Rambler
- Simon Peyton Jones' Implementation of Functional Programming Languages for more about this and related terms.
- Memoising constant applicative forms