Defines functional interfaces for 0-ary, 1-ary and 2-ary functions: F0
, F1
, F2
.
Operations on functions: are defined as "functions" in the Core
interface:
identity
compose
memoize
partial
The constantly
"function" is defined in each functional interface. Its variants couldn't be defined in the Core
interface since its return type varies but its argument list is the same for each arity.
The library operates on "functions", where a function can be any of:
- any lambda:
- the result of any lambda expression or statement
- any method reference
- any object implementing one of the functional interfaces defined in FJ:
F0
,F1
,F2
Because the library functions are themselves available as (static) method references, any function from the library itself, or returned by a function from the library, may be operated upon further.
This works because functions in the library are defined in terms of F0
, F1
, F2
and the functions in the library are closed over those three interfaces plus lambdas.
Unfortunately, due to the way Java 8 works, the library will not work with objects implementing other functional interfaces (not defined in the library). This issue is not unique to FJ--java.util.function
has the same issue. See LambdaTest
for details. This constitutes a barrier to interoperability between functional Java libraries.
From CoreTest
:
Compose two functions:
final F1<Double, Integer> f = compose(x->x + 1.0, x->x * 2);
assertThat(f.apply(3), is(7.0));
Partial application:
final F1<Integer,Integer> f = partial( (x, y)->x + y, 2);
assertThat(f.apply(3), is(5));
Even more partial application:
final F0<Integer> f = partial( (x, y)->x + y, 3, 2);
assertThat(f.apply(), is(5));
And the obiligatory (recursive) fibonacci function with memoization optimization from FibTest.java
:
public class FibTest {
static final F1<Integer,Integer> m_fib = memoize(FibTest::fib);
static int fib(final int n) {
switch(n) {
case 0: case 1: return n;
default: return m_fib.apply(n-1) + m_fib.apply(n-2);
}
}
@Test
public void fibTest() {
// fib(50)
assertThat(m_fib.apply(46),is(1836311903));
}
}
For more usage examples, see CoreTest
, ClosedTest
.
Java 8 introduced lambdas, functional interfaces, method references and a library of ready-made functional interfaces in java.util.function
. That package can be hard to understand all in one go. And while it may represent the definitive intent of the language designers with respect to functional programming, variations on that approach may prove useful, educational, or at least interesting.
FJ is just such a variation. Two constraints differentiate FJ from java.util.function
and other Java functional programming libraries:
- prefer a functional approach over an object-oriented approach
- be tiny: implement only enough to illustrate feasibility
If FJ is small, it might serve as a tool for understanding other functional programming libraries like java.util.function
.
FJ is small. A recent code count:
lines of code | |
---|---|
implementation | 74 |
test (happy path only) | 73 |
The preference for a functional approach is exemplified in FP's complete avoidance of instance methods. Whereas java.util.function
exposes all its behavior as instance methods, FP exposes all its behavior as static methods.
This approach has yielded some interesting results in FJ so far:
- concise interface: fewer concepts ("verbs") for users to learn
- cohesive implementation of each concept
- functional invocation syntax of
Core
"functions" via static import
This table summarizes the concision (of use) and cohesiveness (of implementation) of the FJ appraoch versus java.util.function
:
concept | Java 8 java.util.function |
FJ |
---|---|---|
identity | identity methods in 5 interfaces |
1 identity method in Core |
functional composition | compose and andThen methods... |
3 compose methods in Core |
...in 4 and 10 interfaces respectively | ||
memoization | no | 3 memoize methods in Core |
partial application | no | 3 partial methods in Core |
constantly function |
no | constantly methods... |
...in 3 interfaces: F0 ,F1 ,F2 |
Users have one compose
concept to learn versus 2 for java.util.function
(compose
and andThen
).
Users have one place to go to learn about compose
versus 14 for java.util.function
. A user (or maintainer) seeking the to understand the implementation reads 3 methods, together in a single file versus 14 methods in 14 separate files.
Users have one place to go to learn about identity
(versus 5 places in java.util.function
). A user (or maintainer) seeking the to understand the implementation reads one method versus 5 methods in 5 separate files.
FJ is less broad than java.util.function
though:
Java 8 | FJ | |
---|---|---|
fns on boxed types | yes | yes |
fn arity 0,1,2 | yes | yes |
fns on unboxed types | yes | no |
Since FJ does not address unboxed types, you won't find equivalents of the various DoubleXXX
, IntXXX
, or LongXXX
functions in FJ. Nor will you find boolean-valued functional interfaces like Predicate
or XXXPredicate
in FJ. All of these, could in concept, be added of course, but that isn't the point of FJ.
Also, whereas java.util.function
parameterizes interfaces on the return value last, FJ parameterizes on the return value first.
Java 8 | FJ |
---|---|
Supplier<R> |
F0<R> |
Function<T,R> |
F1<R,T> |
BiFunction<T,U,R> |
F2<R,T,U> |
And actually, FJ uses A
,B
,C
as the formal type parameters:
formal type parameter | semantic meaning |
---|---|
A |
return type |
B |
first argument type |
C |
second argument type |
com.thoughtpropulsion.fj
<dependency>
<groupId>com.thoughtpropulsion</groupId>
<artifactId>FJ</artifactId>
<version>1.0-SNAPSHOT</version>
<dependency>
Java syntax for invoking a static method is concise. By statically importing Core
, compose()
can be invoked like this:
identity(...)
Unfortunately, Java syntax for acquiring a method reference is not only verbose, but it requires the class name:
compose( Core::identity,...)
In Lisps, these two kinds of use are unified syntactically:
(identity ...)
(compose identity ...)
Java type erasure means you can't cast expressions to parameterized types. That, coupled with Java's limited type inference means that this statement:
final F1<Integer,Integer> h = compose(Core::memoize,Core::partial).apply((x,y)->{++counter; return x*y;}, 4);
Results in this compile error:
Error:(25, 40) java: incompatible types: cannot infer type-variable(s) A,B
(argument mismatch; java.lang.Object cannot be converted to com.thoughtpropulsion.fj.F1<A,B>)
Since there was no way to cast the result of compose()
, the statement had to be split in two:
final F2<F1<Integer,Integer>,F2<Integer,Integer,Integer>,Integer> g = compose(Core::memoize,Core::partial);
final F1<Integer,Integer> h = g.apply((x,y)->{++counter; return x*y;}, 4);
The LambdaTest
captures some assumptions about how lambdas, functional interfaces, and method references, work.
It documents what I think is the most mysterious part, namely, that invocation of a lambda (or method reference) through a functional interface actually works, even though neither defines the (named) method of interest!
That would have seemed a lot less mysterious if Java provided a syntax for invoking through a functional interface, without specifying the method name. Proposals were discussed on the lambda-dev mailing list back in 2012 but none were adopted. See f(x) syntax sugar in Lambda
Also documented, is the unfortunate fact that not all functional interfaces are created equal. While you can pass any lambda to a method expecting a functional interface, you can't pass any object implementing any functional interface. If you pass a non-lambda, it has to implement the particular functional interface the method is declared to accept. This is a barrier to interoperability between functional libraries.
The original thinking that led to the FJ library is captured in the F00
interface and F00Test
.