Inside 245-5D

Existential Pontification and Generalized Abstract Digressions

HIW’18: Let’s Go Mainstream with Eta!

My name is Rahul Muttineni, CTO of TypeLead, working on building services around a language named Eta. To get started, I'll give an overview of how the project started, and where it is now.

It started as a HSOC project. It was called GHCVM; back then we had plans of making it both on JVM and CLR... we don't think about CLR anymore. I was mentored by Edward Kmett. We got pretty good response on this, so Jo and I decided to take the risk and work on this full time.

Big thanks to the GHC team, really good work. We've worked with the codebase for two years, and the more and more we work with it, we see how much awesome stuff there is. I've learned a lot by working with the code.

What is Eta? Eta is a fork of GHC. During the GSOC project, it started off as a Haskell program that used the GHC API. Midway in the program, I found that there were certain things that I wanted to do that I couldn't do, and I spent 3-4 days setting up a fork. I'll talk about what those limitations are. Like Haskell, it's a ... language, but the key difference is that it runs on the JVM. That is its own set of challenges, primarily with respect to tail calls. The nice thing about Eta is that it runs on the JVM, and it can run a good chunk of projects just like that. lens... recently, in the last month, we got Yesod working... it's in good shape. The next really great type of Eta is the strongly typed FFI. That works really well with the subtyping in JVM. A good chunk of the talk is about how we got that working. One of the main focuses of Eta is to be focused on industrial use. GHC is focused on industrial use, and research, both. There's a tension between the two... the nice thing we have for Eta is we don't have to face that tension; it's easy to make decisions on how to add new features, because will it help companies? If it is yes, otherwise we don't. (SPJ: That can be a hard question to answer!)

Haskell: Avoid success at all costs. We're not going to sacrifice core principles of language for benefit. Pursue success, at minimum cost. We want to make it successful as much as possible, but we want to make as little sacrifice as possible. That will be a little tricky...

What is Eta? What language features does it support? It started off as a fork of GHC 7.10.3. All extensions that work there, work with Eta as well. The only thing was TemplateHaskell and QuasiQuotes didn't work for a long time. We got it working 3-4 months ago. Biggest change is JavaFFI. GHC 7.10.3 is MINUS C FFI. We could have supported it: Java has JNI, but we tried to avoid it because we didn't want to make platform specific bindings to all the libbraries.

Joe backported a bunch of GHC 8 features: StrictData, ApplicativeDo, OverloadedLabels. Backpack was got recently. There's a very particular reason we had to do it: it has to do with the fact that we don't have green threads by default, and we wanted to give the user a choice of threaded runtime versus blocking runtime.

The compiler? It's a fork of GHC, so all the compiler passes are the same. We just chopped off everything after STG; e.g., C-- is gone. We generate bytecode from STG. We don't do any optimizations right now, and won't need to for some fine. We don't have to because in JVM, it's JIT compiled, so we don't have to optimize as much since JVM will remove a lot of the code that's not used anyway. And the driver: GHC generates object files... we decided to use JAR files. They're just zip files that package up a bunch of class files that store Java bytecodes. We also added one more mode for Uberjars. These are JAR files that are packaged up into one giant package.

I'll talk a little bit about how we implemented the REPL; template haskell. It works through the external-interpreter architecture. In GHC that's called iserv: the process, what it does, is handles running the code. So the compiler will still do the typechecking and everything, but once it's done with all that stuff, GHC will generate, a specific bytecode set, for interpreting Haskell efficiently. Because we already generated JVM bytecodes. We didn't need that custom bytecode set; we just compile with optimizations off; that gives us JVM bytecodes, then we send it to the external process, load it up, and execute them. Implementing the REPL is pretty easy how to get all this working together. JVM has a mechanism called classloading, which is very flexible. You can download bytecodes from the network, get code an runtime. Once you load the class, it's statically compiled code, it's optimized the same, etc.

The build tool we use is Etlas. We didn't want to move too far off of GHC, we stuck with Cabal. At the point we started using it, we forked off of Cabal 2.0. Main difference is that it lets you manage Eta versions. Etlas is almost like Stack, but it's much much closer to Cabal. We took the nice features of Stack and added them to Cabal. The other thing is that it does patch management. What we've been finding as we add more features and backport, Eta is not exactly GHC 7.10, nor is it GHC 8.0, it's a weird intermediate state, so certain packages that won't exactly compile without small changes, so we needed some system to apply those changes before we actually run the build. So we setup a GitHub repository that stores all the patch files. What etlas will do, it will get you the most recent set of patches. Then if you install a package, lens or something, it will download lens, apply the patch, and then it will build. Just recently, we were using base 4.8, and recently we upgraded to base 4.11. But we couldn't update to the new Generics infrastructure, because it slowed down compile times. So there were a bunch of packages that would check if they were GHC 8... and then use new generics. So we had to do a bunch of patching for that. But that's the kind of stuff we have to deal with.

The title of this talk is lets go mainstream with eta. I want to take a moment and say, what does that mean? "The ideas, attitudes, or activities that are shared by most people and regarded as normal or conventional." So at what point does a programming language become consdiered normal or conventional? It has to be used a big company, solve a big real world problem, and people have to believe it works. That's a very complicated question, multifaceted, one part of that answer is, it should make it easier to solve real world problems easier than the status quo. Take for example PHP. PHP came out when there was nothing better to program dynamic web applications. It had just the minimum features required to make it useful to build these. Now everyone here is asking the question: Haskell clearly solves a lot of problems better than the status quo. So why isn't it moving forward? That's a big question, I'm going to talk about how we're approaching it.

The strategy we're using internally, is we put on a "Big Company Hat"; we pretend we're a big company with a lot of employees, millions or billions of lines, and try to figure out what problems they'll face. Some problems are crazy long build times, when trying to build huge software; dynamic where you have to make sure junior developers get up to speed... etc. That's couple to get this conversation started.

After thinking about this a long time, we boiled it down to three basic principles, how we will develop Eta.

1. User Experience
2. Performance
3. Safety

User Experience is mainly, an emotional thing, how you feel when you use Eta technology, how you interact with it, what you feel when you get an error, psychologically. When something has good UX, we feel good. That's a very subjective thing, it can vary between different people, we have to figure out a way to standardize / make it universal. Something we forget as software and tool developers, the person developing the software is human. If they get errors persistently over time, they'll get frustrated. Machines will do what you tell them over and over again.

So what have we done in Eta to concern? We've done something very recently; it's not in master yet. Jo and I spent a week refactoring the codebase to refactor the error reporting part of the typechecker. It stores a list of strings; internally in GHC, there's a pretty printed data type, a list of those. The problem is we can't do postprocessing on that. So, what Jo did was made a giant data type with three hundred data constructors, one for every error message in GHC. That refactor to a week (SPJ: only a week?!) How it is now, it's decoupled, now you have, instead of storing in the typechecking monad, storing strings, you store a data type that stores the relevant data to print out that error message. And then at the final point, you can traverse the data type; based on the presence of other errors, you can decide what to do. Now it's pattern matching on certain error patterns and reporting them nicely. This is one example. We talked about simple errors: refactoring, adding an argument, changing the type, that's one of the most common errors you'll get working with Haskell. So we focused on those first. This shows an example of a type error... 'checker', it's an IO action.

GHC would tell you, couldn't match Int -> IO () with IO (). The problem is, for people who don't know how the typechecker works, they won't be able to understand what the typechecker is doing: going argument by argument. Because of the refactor we've done, it was easy to pattern match on this particular case, and say, hey, if the user forgot to put an argument, you can print out an error message of this form. You print an argument is missing, you highlight. (SM: You might have been missing the first argument, in this case!) That's true. It's tricky; sometimes the suggestion you give, might not. We don't tell people what they did exactly wrong, because we don't know. This is not a perfect thing, but we try to give the best suggestion that we can. And an important feature of this, most of how we decdied this layout, we studied languages like Elm and Purescript, which have done good work in this error. PureScript and Elm both, what they do, for a certain type of error, and you're not sure what to do... e.g., our info is not complete, they can go to a particular link and see other things that could have happened. So we don't have to flood the user with every suggestion, we just have to show the user what probably is the cause for it. And if it's a tricky case, not what we posted, in the link, we'll have the case as well.

(BG: There's other information that might be relevant; expanding type synonyms, etc. Do you have this info?) We're still thinking about that. Probably we'll have extra flags and stuff. Eventually, we'll have a mode that prints out JSON for IDEs, then it's easier to parse on the IDE side. (BG: Incidentally, there's a ticket, a student working with Richard, trying to figure out smoething similar).

Another aspect of UX is we added the REPL. Tried to simplify the entry point, try to make it easy. You want types, kinds, and where to find out more information. This is a statically typed language: you always hhave to be thinking about types. So we :set +t: always print out the types when you print things. One more thing, one of the former Scala engineers, has been learning Haskell, and he made a critique of one aspect of the REPL experience. f is a function of two argumets. In a second statement of the REPL, I applied 1. Find instance, show instance, for a goes to a. He said that... no show instance found, just say that this is a function, and you can't print it. That's a change we did. This was very easy for us to do.

Performance: it can mean many things. We're talking about fast developer feedback loop. Compile time and develop time, reducing that feedback loop. Some work we've done in this direction is reproducible builds. As of now, we have bit-for-bit reproducibility in Eta. That amounted to... nikita already did lots of work on reproducibility, he made HAskell interface reproducible; but the last mile of bit for bit is hard, there's many places. For our code generator, it was a lot simpler, we didn't have to do as much. It was 20 lines of code to make it deterministic. The main source of nondeterminism in GHC is the Unique data type, that changes between different runs depending on environment. What we did, was we added a counter. We used to print the uniques in the Java class name; that will make it nondeterministic. So we made a counter: the order in which the bindings make it to STG is the same.

GHCi is known to take up lots of memory, esp. with IDE. Simon Marlow has a bunch of fixes to that; we also backported those.

Another aspect of performance is the actual runtime performance. We're on the JVM, that puts us at a huge disadvantage. We don't have control over many things. The runtime system... this is Java. It's OO, so the runtime system is implemented in Java. We setup a hierarchy for values, that are defined in Eta. We have Closure, it's a class, parent class of all values, thunks, WNF. The Closure class has two methods. evaluate, evaluate to WHNF, and enter will actually enter... it's similar to GHC runtime system. The initial version was modeled exactly after GHC, except for tail calls. The terminology is similar. It's primarily used when you do the body of function. The main subclasses of Closure are Thunk and Value. Value will be the parent class, of things like functions, partiallly applied functions, and data constructors. Thunk will be the superclass of things like CAFs, single entry thunks, and updatable thunks. CAFs don't have free variables, so there's a special case for that, and you create a blackholing entry every time, to avoid two threads evaluating the same thunk. UpdatableThunk pushes an update frame, when it's finished evaluating, it will update the thunk to point to the newly computed value. And SingleEntryThunk, they're evaluated only once, so you can just evaluate it directly without pushing an update frame. This terminology is borrowed from GHC as well.

VAlues: DataCon, Function and PAPs. In the early days, and even now, every function call that was a tail call, is just a method call. This is the only way to make it remotely efficient. (More on stack soon). For static tail recursive calls: singly recursive or mutually recursive, they get compiled to loops. In most cases, they get a nice tight loop. In the mutual case, what will happen is, we collect all of the SCC, and we make one giant method that goes into a loop. Let's say you're in the even/odd example, what will happen is, when even calls odd, there's a variable called target, an integer. Even will be assigned 0, odd is assigned 1, so then you set 1 and restart. (BG: Do you always have unfoldings available for functions you compiled?) This is mutually recursive functions defined in the same module. (SPJ: They might have very different type arguments.) We cat all the arguments into one. The main problem with this argument, is parsers generated with Happy and Alex, we hit limits. (BG: Crash?) Not stack blowup. JVM has method size limit, so you can only have 65000 bytecodes. That's Eta compiled with itself. That's the only thing that's preventing us from using Eta with Eta. But all you need to do is split method into smaller chunks.

So how do we handle tail calls? When we know it , tail recursive, let's say you don't. Let's say you're using CPS. It's so common in Haskell, any fast parser uses CPS. In early days, Aeson would just blow the stack, it was pretty bad. So, we explored trampolining by default, and it was just awful, it was slow, super slow. What we did is turn it off, and let stack blow up. We found a better solution. The JVM has... the only way to unwind the stack is throwing an exception, or returning, and keep on returning until you return all the way down. It turns out, with exceptions, you can turn off the feature that captures the stack trace: that's the most expensive part of an exception. So we have a general exception. So this trampoline mechanism is optional. So, what we do, we have a function 'trampoline :: a -> a', runtime primitive, what it does is activates a boolean in the context which says, I'm going to trampoline now, and it activates a codepath that turns a counter, and once you reach a certain number, which is configurable, it will unwind the stack, and then continue where it needed to go. Our limit is 400, and then we unwind. It used to be in 1000s, but with Happy and Alex, we needed a smaller number. (BG: Inside that context, how much does it cost? But observably, it's faster. A couple months ago, we got PureScript to work in Eta, and it wasn't bad by default?) (SPJ: So you could turn it on by default: all you're doing is counting.) The counting is how we know how big the stack is. In your main function, you could call trampolineIO, and trampoline your whole program. (SPJ: Maybe it's low overhead, and you can do it all the time.) If it's low, we will do it. (How do you resume? Once you raise the exception, what do you store?) The counter happens at the entry point, and it's guarded bby the boolean. So, that, if the limit is exceeded, it will call another function that takes the context. So we store all the arguments in a context variable that gets passed to every eta function. We stash all the arguments into a function that has the state, then wjhen it unwinds, marked by this function, it will call that, with that function and those arguments.

As I mentioned, it's guarded by a boolean. JVM has an optimization, where it observes the boolean is true for a lot of times, it won't even compile that branch in the native code. So if you don't use trampolining, it doesn't affect you at all; the code for the counter will just not be there.

One nice thing I like about Eta is that you actually get stack traces for exceptions. This is because, to get good perf for Eta, you have to implement most primitives on JVM stack. This is a sample stack. You have a schedule loop, and you hare evaluting some IO acttion. applyV/applyN, these are partial applications. Execute an IO action. And another nice part, we've tried to encode it close to the original name. So you can tell this fucntion call happened in statistics.Regression, rnfAll. If you see, you notice there are line numbers. This is not perfect, and we can definitely make it better later... GHC gives you a lot of debugging info at STG time, but because the JVM doesn't have much flexibility, we can only attach one line number to code, so we have to discard all that info. This will get better; we'll stash that debug information in the classfile itself, and then access it and render a better stack trace. (BG: This is Source Notes?) Yeah.

Concurrency: One nice part is, it's nice or not. If you're evaluating a long chain of thunks, you're going to blow the stack. This happily coincides with GHC also having a space leak. Neil Mitchell wrote a blog post about how to detect space leaks: restrict stack size and then figure out which thunk was being evaluated. If you see a stack trace like this, and you see a huge chain of evaluates, in a long chain, you probably have a space leak.

How do I do interop? The way we did interop was, made a thing called the Java monad. IT's supposed to give you the experience of programming JAva. The basic implementation is inspired from IO monad. Object# c is "this", the object that is being threaded through. Because of this encoding, you get the Java experience: you can call dot on the Java object. It's almost like working with Java inside. The argument is called... that's the type constructor that forced us to fork, instead of use the API. You can't declare primitive types in the API. And we had to introduce a new low level representation. Declare wrapper types, wrapping the iterable interface in Java. We've stolen better syntax, which were type applications... resolve it somehow. I'm declaring an Eta type that wraps a JAva type, @java.lang.Iterable.

You use the java function to run the Java monad. All of these have to be imported. newArrayList, newInteger, but we brought some combinators, that let you call methods. It owrked out with the monad. This is sample code that does the same thing as Java code. it just uses standard monadic combinators. If it's a fixed c, it's an instance.

You can use Eta as a better JAva, with referential transparency! Unlike Kotlin or Scala.

How do we handle subtyping? We define uilt in type families. We have a typeclass named extends. Any time you declare a function that takes a given class and any subtype of that class, you can, instead of actually subtyping, we do it with constraints. Extends' takes the info from Inherits and figures it out. You can use the dot operator on anything that is a subclass of Iterator. We had to extend the typechecker just a little bit: a lot of times the type gets stuck in the form Extends' (List JSTring) (List a) where a is unconstrained.

Imports are tiresome, so we're setting up direct Java Interop; actually use JAva reflection to get info class files, and generate imports. "import java java.lang.Math" works, but doesn't scale. Biggest priority for the rest of the year is Java interop, really good IDE support, documentation, language extensions: UnboxedSums, TypeApplications, DerivingVia, QuantifiedConstraints. We have some new language extensions in mind, AnonymousRecords, RowTypePolymorphism... We'll see how that goes.

I was thinking about ways... we work on the same codebase, how to collaborate? We're interested in compile performance, support for unbboxed sums. Worker wrapper has some glitch, and no one got around to fixing it. At some point, maybe not any time soon, that and mutable fields. Pretty important for us. (BG: Do unboxed sums get a lot of usage? Why unboxed sums? Does Eta code make a lot of use?) No. But a lot of people on JVM are annoyed that Maybe is boxed all the time. But if you have unboxed sums, you can represent it as null. (SPJ: Or you can say, just box it, and you won't notice it. If it's fast enough all the time, focus on what's going to make a difference.)

Q: Did you consider using Graal (it's a new virtual machine that supports partial evaluation and partial escape analysis, good for functional languages)?

A: We have looked into it, it's not completely there yet to use, and we're not sure if it's something we can invest time with. We're keeping up with it. (BG: But you lose the JVM!) That's what's preventing us from going there. Maybe if it gets integrated into a mainline VN we might look at it. (Mainline Java is planning to integrate Graal)

Q: (SM) Are you keeping the fork up to date with master GHC?

A: One thing that is out of bounds for us, and for a long time, is all the dependent Haskell work. Everything else, we keep up. If there's any nice bugfixes... (SM: So you're selectively backporting).

Q: (BG) Have you considered unforking.

A: Not yet, no.