Frege: A Haskell-Like Language for the JVM
infoq.com
infoq.com
There is a significant portion of the Scala community which desperately wants Scala to be the Haskell they can get their bosses buy in for. IMO the result is unsatisfactory, both for those people and the Scala community as a whole.
If a straight Haskell variant in the JVM ecosystem were to "take off" then Scala could be its own thing.
Most scala users have no f-ing clue what a monad is :)
I was making a jest to say that most scala programmers don't know anything about FP and just use it for the type inference and stick to OO practices, vars, etc.
They just don't understand / have much interest in the intermediate/advanced parts. I've seen lots of beginners take to for comprehensions, maps, filters etc pretty quickly.
The problem with Scala is that at an advanced level the code can become pretty unreadable.
I think so. By the time I heard of Kotlin, I mentally lumped it in the basket of "just another JVM language". I mean, there are so many now, and you already had Groovy, Clojure, Scala, JRuby, and Jython pretty well established, along with a couple of dozen other niche players. Then Kotlin comes along... I don't know about most people, but I never heard anything about Kotlin that was so compelling that I felt the need to go mess with it. I mean, why what instead of Nice, or Fantom, or Gosu, or Beanshell, or Mira, etc., etc...
Same thing for Ceylon. It looks like just another "also ran" to me. If the people behind these languages want people to use them, they are going to have to work hard to get the word out about whatever advantages (purported or real) they have.
As a result, it is also the non-Java JVM language with the best IDE support, best Android support and best build-tool support.
Truth be told, I don't really find that very compelling. I mean, do programming languages always need "big company" support to be successful? I don't know. So far Scala, Clojure and Groovy have done pretty well, with varying degrees of commercial support.
As a result, it is also the non-Java JVM language with the best IDE support, best Android support and best build-tool support.
I guess, but here's the thing... the languages I use now (mainly Groovy and Java) have at least good enough IDE support, Android support (ok, I don't really care about that) and build-tool support. So even if Kotlin is incrementally better, or hell, even dramatically better, that stuff doesn't do a lot to sway me to Kotlin.
That said, like most geeks, I like dabbling with new languages, and I am sure I'll try out Kotlin (and Ceylon) at some point. And maybe then one or the other will "wow" me. But for now, it isn't a high priority. And I expect a lot of other people are approaching it with a similar mindset.
History shows that the answer to that is a resounding yes. Aside from a couple of scripting languages (BASIC and Python), there have been exactly zero successful mainstream languages without a large company behind them. Scala, Clojure and Groovy are doing fine, but their adoption is one or two orders of magnitude less than the mainstream languages (Java, C, C++, C#).
I'm not necessarily the biggest fan of the TIOBE index, but in this case I'll cite it, as it is "close enough" I think. Ruby is #13 right now, which isn't bad. And while I think TIOBE has some warts, I think it's safe to say that anything in the top 20 is fairly successful, and anything in the top 50 is "successful" to a certain degree.
http://www.tiobe.com/index.php/content/paperinfo/tpci/index....
You do know VMware/Pivotal pulled their support for Groovy in March this year and no-one else stepped in, don't you?
So? I'm not making a claim that a language needs adoption on that scale to be considered "successful". To me, Ruby, PHP, Python, Groovy, Scala, Clojure, etc. are all very much "successful". The languages I think of as being less so, would be things like Nice, Fantom, Frege, Forth, Modula-3, Rebol, etc.
I do, but I also know that Groovy is becoming an ASF project[1]. So we'll see how it goes with volunteer support, and perhaps a few paid people here and there from companies that use Groovy.
Groovy has lingering problems in migrating to builds.apache.org, see [1]. Some of the Groovy despots from Codehaus times (I wouldn't know which ones specifically) are keeping some computing machinery physically separated from the Apache infrastructure to run their own build processes, going against Apache guidelines. Some of them also have control over the groovy-lang.org domain (again, I don't know who). I suspect Groovy won't ever become an ASF project, but instead just sit it out in the incubation system until the former Codehaus Groovy despots get a better deal where they don't have to share their control democratically with the ASF. They managed to keep Groovy in the Java Community Process for 9 years before being booted out, all the while using the JSR-241 to promote themselves, so they'd think nothing of leeching on Apache's incubator for just as long before stirring up conflict later on to get booted out when it suits them.
BTW, that link [1] to Nabble's Groovy mailing list archive was recently redirected to an embedded view within the groovy-lang.org website where they can collect IP addresses, and all that implies -- another indictment of the Groovy despots longstanding intention to control and instead of sharing.
[1] http://groovy.329449.n5.nabble.com/Jenkins-Groovy-Apache-etc...
Scala has about 2% mind share on the JVM after ten years in existence, I would say that not only is it not winning but its time has passed.
Ceylon is one year old, Kotlin is not even out yet, there is plenty of time for either of these two to gain some solid mind share.
Or maybe not. Maybe Java is going to reign supreme for quite a while. Whatever the replacement of Java will eventually be, I'm pretty sure it won't be Scala.
There just doesn't seem to be a good point in using Kotlin, as you can tell from how fast the Kotlin proponent in this thread is changing the frame of reference.
- Higher-kinded types? Better compare with Haskell!
- Compilation speed? Let's pick Java, the language with the least useful typesystem!
- Popularity? Let's compare Scala with Java, but compare Kotlin to Scala!
- Commercial backing? Let's conveniently ignore that the language lives by the cross-subsidization of JetBrain's IDE business, instead of relying on the success of their language.
Point is, there is nothing that Kotlin does substantially better that would give it a niche between Java and Scala.
- Non-incremental compilation speed is not that much faster compared to Scala, and substantially slower than Java.
- Most of the things regularly used in Scala cannot be expressed in Kotlin, it only makes Java idioms slightly nicer to write down.
- Like in Scala, IDE support is ongoing work.
- Compared to both Java and Scala, the ecosystem isn't there.
Compared to Scala, Kotlin makes you pay 80% of the cost for 20% of the benefits.
I predict that future Java releases will keep cannibalizing Kotlin from the bottom, while Kotlin will fail to attract developers from more expressive languages.
Depends on how we define FP. Until 2010, when Haskell started to emerge as a HN fad, people were OK to define FP as what LISP programmers do, and didn't demand FP programmers know what a modad is, type theory or even use "immutability" everywhere...
If you read discussions from 2000-2005 for example, very few people define FP as "what Haskell does".
Once I finally saw some of the things you can do with a monad rather than those useless examples based around Maybe, I had an 'aha!' moment as everything went click.
No sodding idea what a monoid is, though.
E:
I shouldn't even call it an understanding of monads. I've read at least 15+ explanations of what a monad is and still don't grok it.
Firstly, saying "`X` is a monad" is the same as saying "class `X` implements the monad interface".
The monad interface is usually defined with `return` and `bind`, but I think it's more instructive to borrow the terminology from JavaScript's `Promise`...
* `Promise#resolve(value)` is the same as `return`, also known as `pure`. It just wraps the given value in a promise.
* `Promise#then(function)` combines both `bind` (when a new Promise is constructed and returned) and the functor method `map` (when a plain value is returned).
To provide a unified monad/functor interface for both `Promise` and `Array` (using `wrap` instead of `resolve):
// `then` provides both `bind` and `map` depending on the return type of the given function:
Promise.prototype.map = Promise.prototype.then
Promise.wrap = Promise.resolve
// f maps elements to arrays of elements:
Array.prototype.then = function (f) {
return [].concat.apply ([], this.map(f))
}
Array.wrap = function (x) {
return [x]
}
EDIT: BTW, I've purposely skipped a few things in this description, it's meant to be illustrative, not definitive...
The operation of a monoid maps from pairs of things to things. So in terms of types:
<a,a> -> a
twin a -> a
f a -> a(I'm not a Haskell expert, I dabble, done CIS194, half of RWH, and I have no idea what you said -- which is a common problem I run into in the Haskell world, lots of super helpful people that have forgotten what its like to not speak their language)
For example, we can combine two integers by adding (+) with 0 as an identity. Or we could combine them by (*) with 1 as an identity.
Lists are a monoid because we can append two lists and the empty list is the identity.
The combining operator has to be associative and the identity has to be an identity for it. (Combining something with the identity gives you that original thing back.)
And that's all. There's just not much structure to them. In fact, there's so little structure that mathematicians don't care much about them. It's an exotic-sounding name for a fairly pedestrian concept.
But they're useful in programming. The main reason is that they're so ubiquitous: so many different things form monoids, often in different ways. This means we have a single interface applicable to almost every domain you can think of which is useful even if the interface doesn't tell us much.
They're also a good fit for parallelism. Because the combining function is associative, we can a bunch of combinations in any order we like, making it easy to spread them out over multiple threads. It very naturally captures the reduce in map reduce.
But mostly it's a convenient abstraction that's minimal and simple enough to pop up everywhere while still being useful.
They are HUGELY useful. One of the most useful is a "union" monoid definition for Maps. In Scala, it's written like this:
implicit def mapUnionMonoid[K, V: Semigroup]: Monoid[Map[K,V]]
So any map with Semigroup values is a monoid (Semigroups are monoids without the zero value). |+| under this definition will combine the Maps, but in hte case of collisions, |+| the colliding values together. This Monoid is the ABSOLUTE KING of aggregation. You can foldMap over lists and produce singleton Maps of the shape you want, and let the monoid instance aggregate for you.
monads is that they're way too low level t
Monads are not that exotic. People present it as "deep math" but for a mathematician for example it's kids stuff at the level you need to understand them for Haskell et al.
It's like saying differential equations are some "deep voodoo math" (only monads are even simpler).
Yes, they're much simpler than differential equations. But the difference is, differential equations are the simplest way to solve some difficult problems -- monads are often the gratuitously complex way to solve simple problems.
Consider the Maybe monad - it encodes optional values in a simple way. The approach before this is to have something called "null" that would wreck your programs. Maybe monad is one of the best improvements to day-to-day programming tasks that I've personally experienced in my lifetime.
But thinking monadically and abstracting monadically is extremely different from what programmers normally learn, for a start because important monads like state and exceptions are built-in features of most programming languages. Seeing that these things have a common pattern, and seeing that it may be worthwhile to abstract this common pattern takes a lot of time.
The mismatch between the utter simplicity of the concept of monads, and the complicated explanations one comes across doesn't help.
James is one of the few people who don't make Monads into a mythical being.
http://james-iry.blogspot.de/2007/09/monads-are-elephants-pa...
Indeed I often program in a way that could be called "locally stateful, globally pure-functional". It's a good approach!
Can you expand on that, please? Because it seems to me that state fulness propagates up through the system. I can't imagine a system that was pure in the large but side-effecting in the small. I'd be interested to hear how that works...
If only there was an effect system that could guarantee purity and at the same time not be in the way (i.e. allow full or Scala-like type inference). Then purity would be guaranteed by types, like it is in Haskell. (N.B. I'm not asking for Haskell's purity by default. I advocate impure as default, with a type guarantting purity.)
And when you use just regular Scala syntax, it's just pretty ugly code overall IMO (and one of the slowest compilers of the 21st century too).
I would love if I could avoid this with Fredge. But, in the end, I would probably be better off using Haskell directly.
As long as you are only bringing things in from Java-land, rather than going the other way though, the wrapping can be fairly minimal.
This is certainly true. You can see Frege as a subset of Haskell 2010 plus some GHC extensions plus the native interface (i.e. the Java FFI).
Unless you really need the JVM, Haskell (i.e. GHC) is the probably the better choice.
Frege is just an offer for the minority that wants pure FP specifically on the JVM. Those people had no real choice until recently, given that CAL is dead, and E. Kmett's "ermine" not yet there.
If you want Java you know where to find it. Likewise with Haskell. Scala draws from both and that's its greatest strength.
If you spend a bit of time there it fades to background noise and #scala is actually a really friendly and helpful place - but "x is hard to do in Scala because Scala [sucks]" is an odd impression to give a newcomer.
http://mmhelloworld.github.io/blog/2013/07/10/frege-hello-ja...
"Wrapping mutable Java stuff in Frege means that everything will end up in an ST monad, which is almost always IO. That means that using Java’s HashSet will force anything that uses that to be in the IO monad itself. So there goes purity and referential transparency, which are some of the best parts of Haskell in the first place."
[0] https://www.reddit.com/r/haskell/comments/3gr7y6/infoq_frege...
[1] http://taylor.fausak.me/2015/06/25/frege-a-jvm-haskell/
EDIT: formatting
This is a bit misleading though I think, as you can actually unwrap and ST unlike an IO - my most common use for ST is actually to use a mutable object in a case where it's inefficient to keep copying immutable (e.g.: large vector or matrix manipulations) - then when you're done the ST monad can be unwrapped and you get the final structure - the entire function remains pure but inside of it, there are calculations that rely on state - but no state enters or leaves the function per say. This would mean using a hash set would be out of the question _for passing around outside of an ST monad_ but you could actually use it internally within functions.
Yes, but this is what the author was saying: you end up writing your code in ST. Remeber, there is no (general) way to freeze a mutable collection that was passed and then use it as if it was immutable. There can be a dozen different threads be busy modifying it.
However, as immutable objects are becoming more mainstream even in Java, there will still plenty of opportunities to make polyglot JVM programs employing pure code. Just not with the JAVA collection classes.
[0] https://github.com/Frege/frege-native-gen/issues/14 [1] https://github.com/Frege/frege/wiki/Differences-between-Freg...
-- int frob(int, char*)
foreign import ccall "frob"
frob :: Int -> CString -> IO Int
refrob :: Int -> CString -> Int
refrob x y = unsafePerformIO (frob x y)
The different syntax should not be a big issue, as Haskell code that uses FFI targeting C is per se not portable.
Better yet, if you know it to be pure, you can give it an appropriate type and just use it in pure code, like:
pure native cos java.lang.Math.cos :: Double -> Double
foreign import jvm "java.lang.Math.cos" cos :: Double -> Double
In fact, all primitive operations, types and so on are defined this way in the Frege Prelude.
But it is also not as relevant as one might think.
Consider how many Haskell programs actually use mutable C data, or export functions that take a foreign ptr to some mutable stuff.
Why should this be generally different in Frege? Useful Java APIs will be wrapped and sanitized through Frege libraries, and that is it then.
You can go directly to Java, just like you can directly go to C in Haskell, but it turns out one rarely actually does this.
This might be the reason that haskell has not enjoyed nearly the adoption of Scala and Clojure.
current :: IO String
current = ...
In addition since Haskell supports multiple case defitions at the top level inline type declarations would be redundant. For example -
current a 0 :: IO Int Int -> Int current 0 b :: IO Int Int -> Int current 0 0 :: IO Int Int -> Int
vs.
current :: IO Int Int -> Int current 0 a = ... current a 0 = ... current 0 0 = ...
current :: IO String : = do
d <- Date.new () -- reads system timer, hence IO
d.toString
myFunctionFoo :: Int -> Int -> Int -> Int
: a b c =
2*a + 3*b + 4*cHow would you make your first variant, work with pattern matching?
I like the second variant better, because it seems to work with pattern matching. It would as well have to work without type declarations, letting Haskell infer the types.
But all in all I prefer the way function definition is implemented now. The record issue is more of a trouble - for it Frege actually has some improvements.
But the records names / TDNR issue, this makes a real difference, I'll definitely take a look at Frege if I'll need functional programming on the JVM ...as Scala just seems too scary for me.
Now, about what I proposed above, I meant the two examples as different cases of the same syntax, newlines should not really make a difference until after the "=". So you'd use the second variant for pattern matching:
myFunctionFoo :: Int -> Int -> Int -> Int
: 1 b c =
1984 - c
: a b c =
2*a + 3*b + 4*c
...and it's probably not worth spoiling the beauty of the ML-style-syntax for this. I admit it, I find it 10x easier to read either Lisps or C-syntax-like languages than MLs, but there is a beauty in the ML way, and math folks seem to love it, so better keep it that way :)
foo = (\a -> \b -> (a+b)*(a-b)) :: Num z => z -> z -> z
Monoid: A + B (|+|)
Functor: M[A] => (A => B) => M[B]
Monad: M[A] => (A => M[B]) => M[B]
Applicative: M[A] => M[A=>B] => M[B]
Kleisli: (A => M[B]) => (B => M[C]) => (A => M[C])Am I correct in thinking that Frege cannot leverage on third-party java libraries? Clojure, another functional language built on JVM, has great support for this (possibly) missing feature.
The article at https://mmhelloworld.github.io/blog/2013/07/10/frege-hello-j... talks about this, but doesn't quite explain the proxy mechanism, it just points out an example.
Not quite true anymore. TO be sure, some (inline) java will still be needed.
Here is an example https://github.com/Frege/frege/blob/master/tests/comp/Issue2... which compiles to a class that implements java.util.Comparator. Instances thereof can be created from Frege by passing a custom comparision function and could be passed to Java.
The example would be useful in cases when you have Frege data in a Java collection and want to sort them. But it is merely there to show the possibilities.
Higher rank types
Somewhere it is said that it has all language features of Haskell 2010. This implies higher kinded types.
But in addition to Haskell 2010, Frege has also higher rank types.
Type inference for types of rank > 2 is undecidable, how does this go together with the claim that Frege has type inference? I'm not a Haskell expert, but I think the
{-# LANGUAGE RankNTypes #-}
Actually, the Frege compiler employs an algorithm described in Simon Peyton-Jones paper "Practical Type Inference for Higher Ranked Types". Ordinary HM types are inferred, and higher ranked types checked.
Regarding the pronounciation, who cares? For example, in Germany, half of the people say "Ay-Bee-Em", the other half pronounce the letters IBM in the german way.
Here is a paragraph from "Funktion und Begriff" (1891):
> Wie nun Funktionen von Gegenständen grundverschieden sind, so sind auch Funktionen, deren Argumente Funktionen sind und sein müssen, grundverschieden von Funktionen, deren Argumente Gegenstände sind und nichts anderes sein können. Diese nenne ich Funktionen erster, jene Funktionen zweiter Stufe.
For non-german speakers: Frege makes a distinction between functions that take things as arguments and functions whose argument are and must be functions. He calls the former ones "first order functions" and the latter ones "second order functions".
Today we call functions whose order is greater one "higher order".
I wonder if Frege's older Begriffsschrift (1879) doesn't already discuss, or at least mention, higher-order functions. After all, in this text Frege explains his new conception of function.
I also wonder if Cantor would have been aware that this is possible.
It shows with :java the produced Java code.
The short answer is that it is not derived from existing Haskell compilers.
If you had a Frege -> Haskell translator, which shouldn't be hard to do, except in edge cases, you could use GHC as a testing oracle for the Frege compiler.
The difficulties are clearly stated in the Hasell wiki here https://wiki.haskell.org/GHC:FAQ#Why_isn.27t_GHC_available_f...
Write a Frege-to-Haskell compiler F2H, and then for each test T you simply compare the output of running Frege( T ) with the output of running GHC( F2H( T ) ). Maybe you have to transform the outputs into a universal format such as ASCII strings, but that should be straightforward. Now you have an oracle for random Frege programs.
main = putStrLn "\n"