Elmord's Magic Valley

Software, lingüística e rock'n'roll. Às vezes em Português, sometimes in English.

Posts com a tag: prog

Type parameters and dynamic types

2020-05-30 17:27 +0100. Tags: comp, prog, pldesign, fenius, in-english

In the previous post, I discussed an idea I had for handling dynamic typing in a primarily statically-typed language. In this post, I intend to first, describe the idea a little better, and second, explain what are the problems with it.

The idea

The basic idea is:

For example, consider a function signature like:

let f[A, B](arg1: Int, arg2: A, arg3: B, arg4): Bool = ...

This declares a function f with two explicit type parameters A and B, and four regular value parameters arg1 to arg4. arg1 is declared with a concrete Int type. arg2 and arg3 are declared as having types passed in as type parameters. arg4 does not have an explicit type, so in effect it behaves as if the function had an extra type parameter C, and arg4 has type C.

When the function is called, the type arguments don't have to be passed explicitly; rather, they will be automatically provided by the types of the expressions used as arguments. So, if I call f(42, "hello", 1.0, True), the compiler will implicitly pass the types Str and Float for A and B, as well as Bool for the implicit type parameter C.

In the body of f, whenever the parameters with generic types are used, the corresponding type parameters can be consulted at run-time to find the approprate methods to call. For example, if arg2.foo() is called, a lookup for the method foo inside A will happen at run-time. This lookup might fail, in which case we would get an exception.

This all looks quite beautiful.

The problem

The problem is when you introduce generic data structures into the picture. Let's consider a generic list type List[T], where T is a type parameter. Now suppose you have a list like [42, "hello", 1.0, True] (which you might have obtained from deserializing a JSON file, for instance). What type can T be? The problem is that, unlike the case for functions, there is one type variable for multiple elements. If all type information must be encoded in the value of the type parameter, there is no way to handle a heterogeneous list like this.

Having a union type here (say, List[Int|Str|Float|Bool]) will not help us, because union types require some way to distinguish which element of the union a given value belongs to, but the premise was for all type information to be carried by the type parameter so you could avoid encoding the type information into the value.

For a different example, consider you want to have a list objects satisfying an interface, e.g., List[JSONSerializable]. Different elements of the list may have different types, and therefore different implementations of the interface, and you would need type information with each individual element to be able to know at run-time where to find the interface implementation for each element.

Could this be worked around? One way would be to have a Dynamic type, whose implementation would be roughly:

record Dynamic(
    T: Type,
    value: T,
)

The Dynamic type contains a value and its type. Note that the type is not declared as a type parameter of Dynamic: it is a member of Dynamic. The implication is that a value like Dynamic(Int, 5) is not of type Dynamic[Int], but simply Dynamic: there is a single Dynamic type container which can hold values of any type and carries all information about the value's type within itself. (I believe this is an existential type, but I honestly don't know enough type theory to be sure.)

Now our heterogeneous list can simply be a List[Dynamic]. The problem is that to use this list, you have to wrap your values into Dynamic records, and unwrap them to use the values. Could it happen implicitly? I'm not really sure. Suppose you have a List[Dynamic] and you want to pass it to a function expecting a List[Int]. We would like this to work, if we want static and dynamic code to run along seamlessly. But this is not really possible, because the elements of a List[Dynamic] and a List[Int] have different representations. You would have to produce a new list of integers from the original one, unwrapping every element of the original list out of its Dynamic container. The same would happen if you wanted to pass a List[Int] to a function expecting a List[Dynamic].

All of this may be workable, but it is a different experience from regular gradual typing where you expect this sort of mixing and matching of static and dynamic code to just work.

[Addendum (2020-05-31): On the other hand, if I had an ahead-of-time statically-typed compiled programming language that allowed me to toss around types like this, including allowing user-defined records like Dynamic, that would be really cool.]

EOF

That's all I have for today, folks. In a future post, I intend to explore how interfaces work in a variety of different languages.

1 comentário / comment

Types and Fenius

2020-05-19 21:35 +0100. Tags: comp, prog, pldesign, fenius, in-english

Hello, fellow readers! In this post, I will try to write down some ideas that have been haunting me about types, methods and namespaces in Fenius.

I should perhaps start with the disclaimer that nothing has really happened in Fenius development since last year. I started rewriting the implementation in Common Lisp recently, but I only got to the parser so far, and the code is still not public. I have no haste in this; life is already complicated enough without one extra thing to feel guilty about finishing, and the world does not have a pressing need for a new programming language either. But I do keep thinking about it, so I expect to keep posting ideas about programming language design here more or less regularly.

So, namespaces

A year ago, I pondered whether to choose noun-centric OO (methods belong to classes, as in most mainstream OO languages) or verb-centric OO (methods are independent entities grouped under generic functions, as in Common Lisp). I ended up choosing noun-centric OO, mostly because classes provide a namespace grouping related methods, so:

This choice has a number of problems, though; it interacts badly with other features I would like to have in Fenius. Consider the following example:

Suppose I have a bunch of classes that I want to be able to serialize to JSON. Some of these classes may be implemented by me, so I can add a to_json() method to them, but others come from third-party code that I cannot change. Even if the language allows me to add new methods to existing classes, I would rather not add a to_json() method to those classes because they might, in the future, decide to implement their own to_json() method, possibly in a different way, and I would be unintentionally overriding the library method which others might depend on.

What I really want is to be able to declare an interface of my own, and implement it in whatever way I want for any class (much like a typeclass in Haskell, or a trait in Rust):

from third_party import Foo

interface JSONSerializable {
    let to_json()
}

implement JSONSerializable for Foo {
    let to_json() = {
         ...
    }
}

In this way, the interface serves as a namespace for to_json(), so that even if Foo implements its own to_json() in the future, it would be distinct from the one I defined in my interface.

The problem is: if I have an object x of type Foo and I call x.to_json(), which to_json() is called?

One way to decide that would be by the declared type of x: if it's declared as Foo, it calls Foo's to_json(), and JSONSerializable's to_json() is not even visible. If it's declared as JSONSerializable, then the interface's method is called. The problem is that Fenius is supposed to be a dynamically-typed language: the declared (static) type of an object should not affect its dynamic behavior. A reference to an object, no matter how it was obtained, should be enough to access all of the object's methods.

Solution 1: Interface wrappers

One way to conciliate things would be to make it so that the interface wraps the implementing object. By this I mean that, if you have an object x of type Foo, you can call JSONSerializable(x) to get another object, of type JSONSerializable, that wraps the original x, and provides the interface's methods.

Moreover, function type declarations can be given the following semantics: if a function f is declared as receiving a parameter x: SomeType, and it's called with an argument v, x will be bound to the result of SomeType.accept(v). For interfaces, the accept method returns an interface wrapper for the given object, if the object belongs to a class implementing the interface. Other classes can define accept in any way they want to implement arbitrary casts. The default implementation for class.accept(v) would be to return v intact if it belongs to class, and raise an exception if it doesn't.

Solution 2: Static typing with dynamic characteristics

Another option is to actually go for static typing, but in a way that still allows dynamic code to co-exist more or less transparently with it.

In this approach, which methods are visible in a given dot expression x.method is determined by the static type of x. One way to see this is that x can have multiple methods, possibly with the same name, and the static type of x acts like a lens filtering a specific subset of those methods.

What happens, then, when you don't declare the type of the variable/parameter? One solution would be implicitly consider those as having the basic Object type, but that would make dynamic code extremely annoying to use. For instance, if x has type Object, you cannot call x+1 because + is not defined for Object.

Another, more interesting solution, is to consider any untyped function parameter as a generic. So, if f(x) is declared without a type for x, this is implicitly equivalent to declaring it as f(x: A), for a type variable A. If this were a purely static solution, this would not solve anything: you still cannot call addition on a generic value. But what if, instead, A is passed as a concrete value, implicitly, to the function? Then our f(x: A) is underlyingly basically f(x: A, A: Type), with A being a type value packaging the known information about A. When I call, for instance, f(5), under the hood the function is called like f(5, Int), where Int packages all there is to know about the Int type, including which methods it supports. Then if f's body calls x+1, this type value can be consulted dynamically to look up for a + method.

Has this been done before? Probably. I still have to do research on this. One potential problem with this is how the underlying interface of generic vs. non-generic functions (in a very different sense of 'generic function' from CLOS!) may differ. This is a problem for functions taking functions as arguments: if your function expects an Int -> Int function as argument and I give it a A -> Int function instead, that should work, but underlyingly an A -> Int takes an extra argument (the A type itself). This is left as an exercise for my future self.

Gradual typing in reverse

One very interesting aspect of this solution is that it's basically the opposite of typical gradual typing implementations: instead of adding static types to a fundamentally dynamic language, this adds dynamic powers to a fundamentally static system. All the gradual typing attempts I have seen so far try to add types to a pre-existing dynamic language, which makes an approach like this one less palatable since one wants to be able to give types to code written in a mostly dynamic style, including standard library functions. But if one is designing a language from scratch, one can design it in a more static-types-friendly way, which would make this approach more feasible.

I wonder if better performance can be achieved in this scenario, since in theory the static parts of the code can happily do their stuff without ever worrying about dynamic code. I also wonder if boxing/unboxing of values when passing them between the dynamic and static parts of the code can be avoided as well, since all the extra typing information can be passed in the type parameter instead. Said research, as always, will require more and abundant funding.

Comentários / Comments

Chez Scheme vs. SBCL: a comparison

2019-11-14 11:06 -0300. Tags: comp, prog, lisp, scheme, in-english

Back at the beginning of the year, when I started working on what would become Fenius (which I haven't worked on for a good while now; sorry about that), I was divided between two languages/platforms for writing the implementation: Chez Scheme (a Scheme implementation) and SBCL (a Common Lisp implementation). I ended up choosing Chez Scheme, since I like Scheme better. After using Chez for a few months now, however, I've been thinking about SBCL again. In this post, I ponder the two options.

Debugging and interactive development

The main reason I've been considering a switch is this: my experience with interactive development with Chez has been less than awesome. The stack traces are uninformative: they don't show the line of code corresponding to each frame (rather, they show the line of code of the entire function, and only after you ask to enter debug mode, inspect the raise continuation, and print the stack frames), and you can't always inspect the values of parameters/local variables. The recommended way to debug seems to be to trace the functions you want to inspect; this will print each call to the function (with arguments) and the return value of each call. But you must do it before executing the function; it won't help you interpret the stack trace of an exception after the fact.

The interaction between Chez and Geiser (an Emacs mode for interactive development with Scheme) often breaks down too: sometimes, trying to tab-complete an identifier will hang Emacs. From my investigation, it seems that what happens is that the Chez process will enter the debugger, but Geiser is unaware of that and keeps waiting for the normal > prompt to appear. Once that happens, it's pretty much stuck forever (you can't tab-complete anymore) until you restart Chez. There is probably a solution to this; I just don't know what it is.

As I have mentioned before, Chez has no concept of running the REPL from inside a module (library in Scheme terminology), which means you can't call the private functions of a module from the REPL. The solution is… not to use modules, or to export everything, or split the code so you can load the module code without the module-wrapping form.

By contrast, SBCL works with SLIME, the Superior Lisp Interaction Mode for Emacs. SLIME lets you navigate the stack trace, see the values of local variables by pressing TAB on the frame, you can press v to jump to the code corresponding to a stack frame (right to the corresponding expression, not just the line), among other features. Common Lisp is more committed to interactive development than Scheme in general, so this point is a clear win for SBCL.

(To be fair, Guile Scheme has pretty decent support for interactive development. However, Guile + Geiser cannot do to stack traces what SBCL + SLIME can.)

Execution speed

In my experience, SBCL and Chez are both pretty fast – not at the "as fast as hand-crafted C" level, but pretty much as fast as I could desire for a dynamically-typed, garbage-collected, safe language. In their default settings, Chez actually often beats SBCL, but SBCL by default generates more debugger-friendly code. With all optimizations enabled, Chez and SBCL seem to be pretty much the same in terms of performance.

One advantage SBCL has is that you can add type annotations to code to make it faster. Be careful with your optimization settings, though: if you compile with (optimize (safety 0)), "[a]ll declarations are believed without assertions", i.e., the compiler will generate code that assumes your types are correct, and will produce undefined behavior (a.k.a. nasal demons) in case it is not.

Startup time and executable size

This one is complicated. In my machine, Chez compiles a "hello world" program to a 837-byte .so file, which takes about 125ms to run – a small but noticeable startup time. A standalone binary compiled with chez-exe weighs in at 2.7MB and takes 83ms – just barely noticeable.

As for SBCL, a "hello world" program compiles to a 228-byte .fasl file, which runs in 15ms, which is really good. The problem is if the file loads libraries. For instance, if I add this to the beginning of the "hello world":

(require 'asdf)        ;; to be able to load ASDF packages
(require 'cl-ppcre)    ;; a popular regex library

…now the script takes 422ms to run, which is very noticeable.

SBCL can also generate standalone executables, which are actually dumps of the whole running SBCL image: you can load all the libraries you want and generate an executable with all of them preloaded. If we do that, we're back to the excellent 15ms startup time – but the executable has 45MB, because it contains a full-fledged SBCL in it (plus libraries). It's a bit of a bummer if you intend to create multiple independent command-line utilities, for example. Also, I guess it's easier to convince people to download a 2.7MB file than a 45MB one when you want them to try out your fancy new application, though that may not be that much of a concern these days. (The binary compresses down to 12MB with gzip, and 7.6MB with xz.)

Another worry I have is memory consumption (which is a concern in cheap VPSes such as the one running this blog, for instance): running a 45MB binary will use at least 45MB of RAM, right? Well, not necessarily. When you run an executable, the system does not really load all of the executable's contents into memory: it maps the code (and other) sections of the executable into memory, but they will actually only be loaded from the disk to RAM as the memory pages are touched by the process. This means that most of those 45MB might never actually take up RAM.

In fact, using GNU time (not the shell time builtin, the one in /usr/bin, package time on Debian) to measure maximum memory usage, the SBCL program uses 19MB of RAM, while the Chez program uses 27MB. So the 45MB SBCL binary is actually more memory-friendly than the 2.7MB Chez one. Who'd guess?

Available libraries

Common Lisp definitely has the edge here, with over a thousand libraries (of varying quality) readily available via Quicklisp. There is no central repository or catalog of Chez (or Scheme) libraries, and there are not many Chez libraries that I'm aware of (although I wish I had learned about Thunderchez earlier).

[Addendum (2019-11-16): @Caonima67521344 on Twitter points out there is the Akku package manager for Chez and other Schemes.]

The language itself

This one is a matter of personal taste, but I just like Scheme better than Common Lisp. I like having a single namespace for functions and variables (which is funny considering I was a big fan of Common Lisp's two namespaces back in the day), and not having to say funcall to call a function stored in a variable. I like false being distinct from the empty list, and for cdr of the empty list to be an error rather than nil. I like Scheme's binding-based modules better than Common Lisp's symbol-based packages (although Chez modules are annoying to use, as I mentioned before; Guile is better in this regard). Common Lisp's case insensitivity by default plus all standard symbols being internally uppercase is a bit annoying too. Scheme has generally more consistent names for things as well. I used to dislike hygienic macros back in the day, but nowadays, having syntax-case available to break hygiene when necessary, I prefer hygienic macros as well.

And yet… Common Lisp and Scheme aren't that different. Most of those things don't have a huge impact in the way I code. (Well, macros are very different, but anyway.) One things that does have an impact is using named let and recursion in Scheme vs. loops in Common Lisp: named let (similar to Clojure's loop/recur) is one of my favorite Scheme features, and I use it all the time. However, it is not difficult to implement named let as a macro in Common Lisp, and if you only care about tail-recursive named let (i.e., Clojure's loop/recur), it's not difficult to implement an efficient version of it in Common Lisp as a macro. Another big difference is call/cc (first class continuations) in Scheme, but I pretty much never use call/cc in my code, except possibly as escape continuations (which are equivalent to Common Lisp's return).

On the flip side, Common Lisp has CLOS (the Common Lisp Object System) in all its glory, with generic functions and class redefinition powers and much more. Guile has GOOPS, which provides many of the same features, but I'm not aware of a good equivalent for Chez.

Conclusion

As is usually the case when comparing programming languages/platforms, none of the options is an absolute winner in all regards. Still, for my use case and for the way I like to program, SBCL looks like a compelling option. I'll have to try it for a while and see how it goes, and tell you all about it in a future post.

5 comentários / comments

Functional record updates in Fenius, and other stories

2019-06-16 17:33 -0300. Tags: comp, prog, pldesign, fenius, in-english

Fenius now has syntax for functional record updates! Records now have a with(field=value, …) method, which allows creating a new record from an existing one with only a few fields changed. For example, if you have a record:

fenius> record Person(name, age)
<class `Person`>
fenius> let p = Person("Hildur", 22)
Person("Hildur", 22)

You can now write:

fenius> p.with(age=23)
Person("Hildur", 23)

to obtain a record just like p but with a different value for the age field. The update is functional in the sense that the p is not mutated; a new record is created instead. This is similar to the with() method in dictionaries.

Another new trick is that now records can have custom printers. Printing is now performed by calling the repr_to_port(port) method, which can be overridden by any class. Fenius doesn't yet have much of an I/O facility, but we can cheat a bit by importing the functions from the host Scheme implementation:

fenius> record Point(x, y)
<class `Point`>
fenius> import chezscheme

# Define a new printing function for points.
fenius> method Point.repr_to_port(port) = {
            chezscheme.fprintf(port, "<~a, ~a>", this.x, this.y)
        }

# Now points print like this:
fenius> Point(1, 2)
<1, 2>

A native I/O API is coming Real Soon Now™.

Comentários / Comments

Questions, exclamations, and binaries

2019-06-03 21:39 -0300. Tags: comp, prog, pldesign, fenius, in-english

I'm a bit tired today, so the post will be short.

ready? go!

In Scheme, it is conventional for procedures returning booleans to have names ending in ? (e.g., string?, even?), and for procedures which mutate their arguments to have names ending in ! (e.g., set-car!, reverse!). This convention has also been adopted by other languages, such as Ruby, Clojure and Elixir.

I like this convention, and I've been thinking of using it in Fenius too. The problem is that ? and ! are currently operator characters. ? does not pose much of a problem: I don't use it for anything right now. !, however, is a bit of a problem: it is part of the != (not-equals) operator. So if you write a!=b, it would be ambiguous whether the ! should be interpreted as part of an identifier a!, or as part of the operator !=. So my options are:

What do you think? Which of these you like best? Do you have other ideas? Feel free to comment.

Binaries available

In other news, I started to make available a precompiled Fenius binary (amd64 Linux), which you can try out without having to install Chez Scheme first. You should be aware that the interpreter is very brittle at this stage, and most error messages are in terms of the underlying implementation rather than something meaningful for the end user, so use it at your own peril. But it does print stack traces in terms of the Fenius code, so it's not all hopeless.

Comentários / Comments

Pattern matching and AST manipulation in Fenius

2019-05-30 19:40 -0300. Tags: comp, prog, pldesign, fenius, in-english

Fenius has pattern matching! This means you can now write code like this:

record Rectangle(width, height)
record Triangle(base, height)
record Circle(radius)

let pi = 355/113    # We don't have float syntax yet :(

let area(shape) = {
    match shape {
        Rectangle(width, height) => width * height
        Triangle(base, height) => base * height / 2
        Circle(radius) =>  pi * radius * radius
    }
}

print(area(Rectangle(4, 5)))
print(area(Triangle(3, 4)))
print(area(Circle(10)))

More importantly, you can now pattern match over ASTs (abstract syntax trees). This is perhaps the most significant addition to Fenius so far. It means that the code for the for macro from this post becomes:

# Transform `for x in items { ... }` into `foreach(items, fn (x) { ... })`.
let for = Macro(fn (ast) {
    match ast {
        ast_match(for _(var) in _(items) _(body)) => {
            ast_gen(foreach(_(items), fn (_(var)) _(body)))
        }
    }
})

This is a huge improvement over manually taking apart the AST and putting a new one together, and it basically makes macros usable.

It still does not handle hygiene: it won't prevent inserted variables from shadowing bindings in the expansion site, and will break if you shadow the AST constructors locally. But that will come later. (The AST constructors will move to their own module eventually, too.)

The _(var) syntax is a bit of a hack. I wanted to use some operator, like ~var or $var, but the problem is that all operators in Fenius can be interpreted as either infix or prefix depending on context, so in for $var would be interpreted as an infix expression for $ var, and you would have to parenthesize everything. One solution to this is to consider some operators (like $) as exclusively prefix. I will think about that.

How does it work?

I spent a good while hitting my head against the whole meta-ness of the ast_match/ast_gen macros. In fact I'm still hitting my head against it even though I have already implemented them. I'll try to explain them here (to you and to myself).

ast_match(x) is a macro that generates a pattern that would match the AST for x. So, for example, ast_match(f(x)) generates a pattern that would match the AST for f(x). Which pattern is that? Well, it's:

Call(_, Identifier(_, `f`), [Identifier(_, `x`)])

That's what you would have to write on the left-hand side of the => in a match clause to match the AST for f(x). (The _ patterns are to discard the location information, which is the first field of every AST node. ast_gen is just like ast_match but does not discard location information.) So far, so good.

But here's the thing: that's not what the macro has to output. That's what you would have to write in the source code. The macro has to output the AST for the pattern. This means that where the pattern has, say, Identifier, the macro actually has to output the AST for that, i.e., Identifier(nil, `Identifier`). And for something like:

Identifier(_, `f`)

i.e., a call to the Identifier constructor, the macro has to output:

Call(nil, Identifier(nil, `Identifier`),
          [Identifier(nil, `_`), Constant(nil, `f`)])

and for the whole AST of f(x), i.e.:

Call(_, Identifier(_, `f`), [Identifier(_, `x`)])

the macro has to output this monstrosity:

Call(nil, Identifier(nil, `Call`),
     [Identifier(nil, `_`),
      Call(nil, Identifier(nil, `Identifier`),
                [Identifier(nil, `_`), Constant(nil, `f`)]),
      Array(nil, [Call(nil, Identifier(nil, `Identifier`),
                            [Identifier(nil, `_`), Constant(nil, `x`)])])])

All of this is to match f(x). It works, is all encapsulated inside the ast_* macros (so the user doesn't have to care about it), and the implementation is not even that much code, but it's shocking how much complexity is behind it.

Could it have been avoided? Perhaps. I could have added a quasiquote pattern of sorts, which would be treated especially by match; when matching quasiquote(ast), the matching would happen against the constructors of ast itself, rather than the code it represents. Then I would have to implement separate logic for quasiquote outside of a pattern (e.g., on the right-hand side). In the end, I think it would require much more code. ast_match/ast_gen actually share all the code (they call the same internal meta-expand function, with a different value for a "keep location information" boolean argument), and requires no special-casing in the match form: from match's perspective, it's just a macro that expands to a pattern. You can write macros that expand to patterns and use them in the left-hand side of match too.

(I think I'll have some observations on how all of this relates/contrasts to Lisp in the future, but I still have not finished digesting them, and I'm tracking down some papers/posts I read some time ago which were relevant to this.)

Missing things

The current pattern syntax has no way of matching against a constant. That is:

match false {
    true => "yea"
    false => "nay"
}

binds true (as a variable) to false and returns "yea". I still haven't found a satisfactory way of distinguishing variables from constants (which are just named by identifiers anyway). Other languages do various things:

One thing that occurred to me is to turn all constructors into calls (i.e., you'd write true() and false(), not only in patterns but everywhere), which would make all patterns unambiguous, but that seems a bit annoying.

Rust's solution seems the least intrusive, but Fenius does not really have a syntactically separate class of "constructors" (as opposed to just variables bound to a constant value), and considering all bound variables as constants in patterns makes patterns too fragile (if you happen to add a global variable – or worse, a new function in the base library – with the same name as a variable currently in use in a pattern, you break the pattern). I'll have to think more about it. Suggestions and comments, as always, are welcome.

Another missing thing is a way to debug patterns: I would like to be able to activate some kind of 'debug mode' for match which showed why a pattern did not match. I think this is feasible, but we'll see in the future.

Comentários / Comments

Partial goals, and other rambles

2019-05-25 22:33 -0300. Tags: comp, prog, pldesign, fenius, ramble, in-english

Designing a programming language is a huge undertaking. The last few days I have been thinking about types and classes and interfaces/traits and I got the feeling I was getting stuck in an analysis paralysis stage again. There is also lots of other questions I have to thinking about, such as: How will I handle mutability, and concurrency, and how those things interact? Can I add methods to classes at runtime? Can I implement interfaces to arbitrary existing types? How do I get a trait implementation 'in scope' / available? How static and dynamic typing will interact? How does this get implemented under the hood? And so on, and so forth…

At the same time, it may seem that not solving those problems undermine the whole point of designing a new language in the first place. If you don't solve the hard problems, why not keep using an existing language?

And that's the recipe for paralysis.

But the thing is, even if you solve part of the problems, you can already have something valuable. If Fenius 0.3 gets to be just a 'better' Scheme with nicer syntax and fewer namespacing problems*, that would be already a language I would like to use.

So what would a Minimum Viable Fenius need?

These would go a long way already. These are all pretty doable and don't involve much hard thinking. The greater problems can be tackled later, after the language is usable for basic tasks.

I would also love to have basic sockets to try to write web stuff in Fenius, but these will have to wait, especially given that Chez does not come with sockets natively. Thunderchez has a socket library; maybe I can use that. But this will come after the items above are solved. I can write CGI apps in it, anyway, or make a wrapper in another language to receive the connections and pass the data via stdin/stdout.

Also, I sometimes wonder if it wouldn't be more profitable to implement Fenius on the top of Guile instead of Chez, since it has more libraries, and now also has a basic JIT compiler. But at the same time, the ultimate goal is to rewrite the implementation in itself and leave the Scheme dependency behind, so maybe it's better not to depend too much on the host ecosystem. Although the Chez compiler is so good, maybe it would make more sense to compile to Scheme and run on the top of Chez rather than compile to native code. If only it did not have the annoying startup time… we'll see in the future!

_____

* Of course, 'better' is subjective; what I want is for Fenius 0.3 to be a better Scheme for me.

Addendum: And by 'better syntax' I don't even mean the parentheses; my main annoyance when programming in Scheme is the syntax for accessors: the constant repeating of type names, like (person-name p) rather than p.name, (vector-ref v i) rather than v[i], (location-line (AST-location ast)) rather than ast.location.line, and so on. This goes hand in hand with the namespacing issue: because classes define a namespace for their methods, you don't have to care about name conflicts so often.

Comentários / Comments

Persistent hashmaps

2019-05-23 21:36 -0300. Tags: comp, prog, pldesign, fenius, in-english

Fenius got dictionaries! Now you can say:

fenius> let d = dict("foo" => 1, "bar" => 2)
dict("foo" => 1, "bar" => 2)
fenius> d["foo"]
1
fenius> d.with("foo" => 3, "quux" => 4)
dict("foo" => 3, "bar" => 2, "quux" => 4)
fenius> d.without("foo")
dict("bar" => 2)

Dictionaries are immutable: the with and without methods return new dictionaries, rather than modifying the original. I plan to add some form of mutable dictionary in the future, but that will probably wait until I figure Fenius's mutability story. (Right now, you can mutate the variables themselves, so you can write d := d.with("foo" => 42), for example.)

Dictionaries are persistent hashmaps, much like the ones in Clojure. Fenius's implementation of hashmaps is simpler than Clojure's, because it does less: beside persitent/immutable hashmaps, Clojure also supports transient hashmaps, a mutable data structure onto which entries can be inserted and removed efficiently, which then can be turned into a persistent hashmap. I want to implement something like this in Fenius at some point. On the flip side, the current Fenius implementation is easier to understand than its Clojure counterpart.

The underlying data structure is based on hash array mapped tries (HAMTs). The standard reference for it is Bagwell (2000). The persistent variant used by Clojure, Fenius and other languages is slightly different, but the basic idea is the same. In this post, I intend to explain how it works.

I will begin by explaining an auxiliary data structure used by the hashmap, which I will call a sparse vector here.

Sparse vectors

Suppose you want to create vectors of a fixed size (say, 8 elements), but you know most of the positions of the vector will be empty much of the time. If you create lots of these vectors in the naive way, you will end up wasting a lot of memory with the empty positions.

23 42

A cleverer idea is to split this information into two parts:

00001010 23 42

(Note that the first position is represented by the lowest bit.) If we know beforehand that the number of positions is limited (say, 32), we can fit the bitmap into a single integer, which makes for a quite compact representation.

That takes care of representing the vectors. But how do we fetch a value from this data structure given its position index i? First, we need to know if the position i is filled or empty. For that, we bitwise-and the bitmap with a mask containing only the ith bit set, i.e., bitmap & (1 << i) (or, equivalently, bitmap & 2i). If the result is non-zero, then the ith bit was set in the bitmap, and the element is present. (If it's not, we return a default empty value.)

Once we know the element is there, we need to find its place in the content vector. To do that, we need to know how many elements are there before the one we want, and skip those. For that:

The result of all this is the number of filled positions before the one we want. Because positions are counted from 0, it also happens to be the position of the wanted element in the content vector (e.g., if there is 1 element before the one we want, the one we want is at position 1 (which is the second position of the vector)). In summary:

def sparsevec_get(sparsevec, position):
    if sparsevec.bitmap & (1 << position):
        actual_position = bit_count(sparsevec.bitmap & ((1 << position) - 1))
        return sparsevec.content[actual_position]
    else:
        return SOME_EMPTY_VALUE

To insert an element at position i into the vector, we first figure if position i is already filled. If it is, we find its actual position in the content vector (using the same formula above) and replace its contents with the new value. If it is currently empty, we figure out how many positions are filled before i (again, same formula), and insert the new element just after that (and update the bitmap setting the ith bit). Analogously, to remove the ith element, we check if it is filled, and if it is, we find its position in the content vector, remove it, and clear the ith bit in the bitmap.

In our real implementation, sparse vectors will be persistent, so updates are performed by creating a new vector that is just like the original except for the added or removed elements.

Enough with sparse vectors. Let's get back to HAMTs.

Hash Array-Mapped Tries

A trie ("retrieval tree") is a tree where the path from the root to the desired element is determined by the symbols of the element's key. More specifically:

A hash function is a function f such that if two elements x and y are equal, then f(x) = f(y). (The opposite is not true: there are typically infinitely many elements x for which f(x) produces the same result.) The result of f(x) is called a hash of x. A hash is typically a finite sequence of bits, or symbols... from an alphabet. You can see where this is going.

A hash trie is a trie indexed by hashes. To find an element with key x in the trie, you first compute the hash f(x), and then consume successive bits (or symbols) from the hash to decide the path from the root of the trie to the desired element.

In principle, this would mean that the depth of the trie would be the same as the length of the hash. For example, if our hashes have 32 bits, we would have to traverse 32 children until we reach the element, which would be pretty wasteful in both storage and time. However, we don't need to use the entire hash: we can use just enough bits to distinguish the element from all other elements in the trie. For example, if the trie has elements x with hash 0010, y with hash 0100, and z with hash 1000, we build a trie like:

                     *
                  0 / \ 1
                   /   \
                  *     z
                0/ \1
                /   \
               x     y

That is: z is the only element whose hash starts with 1, so we can store it right below the root as the child labelled 1. Both x and y have hashes beginning with 0, so we create a subtree under the root's 0 for them; the second bit of the hash is 0 for x and 1 for y, so we store them below the subtree under the appropriate label. We don't need to go any deeper to disambiguate them.

To insert an element in the tree, we compute its hash and start traversing the tree until either:

If our hashes have a finite number of bits, it may happen that two distinct elements end up having the same hash. There are two ways of handling this problem:

The only problem with this scheme is that our trees can still get pretty deep as we add more elements to them. For example, if we add 1024 elements (210) to the tree, we need at least 10 bits of hash to distinguish them, which means we need to go at least 10 levels deep in the tree to find them; the deeper the tree is, the slower it is to find an element in it. We can reduce the depth if, instead of branching on single bits of the hash, we use groups of bits of a fixed size, say, 5 bits. Then instead of each node having 2 children, each node has 25 = 32 children, labelled {00000, 00001, ..., 11110, 11111}, and we consume 5 bits of the hash for each level we traverse. Now a tree of 210 elements will typically be 2 levels deep rather than 10, which makes it much faster to traverse.

The only problem with this scheme is that each node now needs to have space for 32 children, even when the children are empty. For example, if I store two elements, x with hash 00000 00001, and y with hash 00000 00100, the root of the tree will be a node with 32 children, of which only the 00000 child will be non-empty. This child will contain a subtree containing x at position 00001, y at position 00100, and all other 30 positions empty. If only we had a way to only store those positions that are actually filled. A sparse vector, if you will...

Congratulations, we have just invented hash array-mapped tries, or HAMTs. A HAMT is a hash trie in which each non-leaf node is a sparse vector, indexed by a fixed-length chunk of bits from the hash. To find an element in the HAMT, we traverse the sparse vectors, consuming successive chunks from the hash, until we either find the element we want, or we consume the entire hash and reach a collision list (in which case we look for the element inside the list), or we reach an empty child (in which case the element is not found). Because the sparse vector only allocates space for the elements actually present, each node is compact, and because each level is indexed by a large-ish chunk of bits, the tree is shallow. Win win.

The sparse vectors are immutable, and so is our tree. To add an element to the tree, we have to change the nodes from the root to the final place of the element in the tree, which means making copies of them with the desired parts changed. But the nodes that are not changed (i.e., those that are not part of the path from the root to the new element) are not copied: the new tree will just point to the same old nodes (which we can do because we know they won't change). So adding an element to the tree does not require making a full copy of it.

Removing an element from a HAMT requires some care. Basically, we have to replace the node where the element is with an empty child. But if the subtree where the element was had only two elements, and after removal is left with only one, that one element takes the place of the whole subtree (you never have a subtree with a single leaf child in a HAMT, because the purpose of a subtree is to disambiguate multiple elements with the same hash prefix; if there is no other element sharing the same prefix with the element, there is no point in having a subtree: the element could have been stored directly in the level above instead).

Miscellaneous notes

When profiling my implementation in a benchmark inserting a million elements in a HAMT, I discovered that most of the time was spent on an auxiliary function I wrote to copy sequences of elements between vectors (when updating sparse vectors). This would probably be more efficient if R6RS (or Chez) had an equivalent of memcpy for vectors. It does have bytevector-copy!, but not a corresponding vector-copy!. Go figure.

R7RS does have a vector-copy!, but I'm using Chez, which is an R6RS implementation. Moreover, R7RS(-small) does not have bitwise operations. But it totally has gcd (greatest common divisor), lcm (least common multiple) and exact-integer-sqrt. I mean, my idea of minimalism is a bit different. Also, it has a timestamp function which is based on TAI instead of UTC and thus requires taking account of leap seconds, except it's not really guaranteed to return accurate time anyway ("in particular, returning Coordinated Universal Time plus a suitable constant might be the best an implementation can do"). Yay. [/rant]

Implementing transient/mutable HAMTs efficiently is a bit more complicated. For it to be efficient, you need to be able to insert new elements in a sparse vector in-place, but you can only do that if you pre-allocate them with more space than they actually need, so you have room for growing. Finding a proper size and grow factor is left as an exercise to the reader.

Comparing performance with Clojure HAMTs is not a very exact benchmark, because the implementations are not in the same language (Clojure's is in Java and Fenius's is in Chez Scheme). In my tests doing 10M insertions, Clojure sometimes did much faster than Chez, sometimes much slower, with times varying between 16s and 48s; the JVM works in mysterious ways. Chez's fastest was not as fast as Clojure, but performance was consistent across runs (around ~35s). Note that this is the time for using the hashmap implementation from ChezScheme, not Fenius; doing the benchmark directly in Fenius would be much slower because the language is currently interpreted and the interpreter is very naive. Note also that in actual Clojure, you would do all the insertions on a transient hashmap, and then turn it into a persistent hashmap after all the insertions, so the benchmark is not very representative of actual Clojure usage.

End of file

That's all for now, folks. I wanted to discuss some other aspects of Fenius dictionaries (such as syntax), but this post is pretty big already. Enjoy your hashmaps!

Comentários / Comments

Modules in Hel (and Chez Scheme)

2019-04-18 22:41 -0300. Tags: comp, prog, pldesign, lisp, scheme, hel, fenius, in-english

Today I implemented a simple mechanism for importing modules in Hel. Basically, you can write:

import foo/bar/baz

and it will look for a file named foo/bar/baz.hel, load it as a module containing the bindings defined in the file, and expose the module as baz to the calling code. So if foo/bar/baz.hel defines a function hello, then after you import foo/bar/baz, you can call the function as baz.hello(). Alternatively, you can import the module with a different name using:

import foo/bar/baz as whatever

and access the bindings like whatever.hello().

That's all there is to it, which means there's a lot of things missing from the system. But I decided it was best to do the simple thing1 and leave the more complex details of the module system for a later phase.

Oh, one more trick: if the module file is not found, Hel tries to import a Scheme module with the given name. So you can actually say import chezscheme and get all the bindings from the host Scheme. (Except some of the names are inacessible, since there is currently no way to use ? or - in Hel identifiers. We'll see how to fix that in the future.

Modularizing the interpreter

Incidentally, I also started an attempt to split the interpreter (a single 1420-line Scheme file) into modules. I still haven't merged those changes into the master repository, and I'm still not sure if it's a good idea. In principle it'd be great for organization. The problem is that the RnRS/Chez module system is annoying to use.

For instance, you have to list all bindings you want to export in the library declaration. This is especially annoying for records. For example, if you declare a record:

(define-record-type Foo (fields x y))

this will generate a record descriptor Foo, a constructor make-Foo, a type predicate Foo?, and two accessors Foo-x and Foo-y. That's all nice and fun, but you have to export each of the generated identifiers individually if you want to use them in other modules.

Another annoyance is that Chez does not seem to provide a mechanism to run the REPL from the environment of the module. You can switch the interaction environment to the exported bindings of a module, but there does not seem to be a way to switch to the environment within the module, to call non-exported functions, etc. The workaround I found was to split all modules into a library definition file (say, utils.sls), containing just:

(library (utils)
  (export binding1 binding2 binding3 ...)
  (import (chezscheme))
  (include "utils.scm"))

and the module code proper, in a separate file utils.scm. In this way, I can load the module code directly in Geiser or in the REPL, outside the module system. Note also that the the library definition file only imports (chezscheme) (so we can use the include form); all other imports are directly in the .scm file, so the .scm file will load properly by itself.

Even so, it is annoying to reload libraries, because you have to reload the users of each library manually too.

A smaller annoyance is that the code takes longer to compile when split into libraries. This is not a problem if you compile before execution, but makes running the code without a separate compilation step (chezscheme --script hel.scm) slower to start.

Yet another annoyance is that in the R6RS library syntax, all definitions must precede all expressions, so you have to move initialization code to the end of the module. [Addendum: this (mis)feature does not seem to be shared by R7RS library syntax. As much as I've learned to appreciate many good aspects of R6RS, it seems to me that library syntax is just better in R7RS.]

In the end, a middle-ground solution may be to avoid R6RS libraries entirely, and just create a single main file which includes the others, all in the same namespace. It's not as elegant, but it makes development easier. [Addendum: one benefit of the .sls/.scm split is that it's easy to switch to the non-library organization by just including all .scm files directly and ignoring the .sls files.]

Todos and remarks

When you write import foo/bar, where to look for the foo/bar.hel file? Currently the interpreter looks in the current directory, but that's far from ideal. A better option would be to search relative to the file where the import occurs. That would be an improvement, but still somewhat annoying: if I have a project with files foo.hel, dir1/bar.hel and dir2/baz.hel, I want to be able to load either of these files individually in the REPL and for each module to be able to import any other module in the project using the same name. What I really want is a notion of a project root to search from. One possibility would be to have a __project__.hel file (or something similar) at the project root. When looking for imports, the implementation tries to find the __project__.hel file up the directory hierarchy. If it's found, the directory where the file is is the project root. If not, the project root is the directory where the importing file is. This is vaguely similar to Python's __main__.py, except there would be only one project file per project (not per directory, which would destroy the idea of a single project root).

There is still no syntax to import individual bindings from a module, i.e., the equivalent of Python's from mod import foo, bar. Maybe we can just use Python's syntax (except (foo, bar) would have to be in parentheses, due to restrictions of Hel's syntax).

There is also no syntax to import all bindings from a module, i.e., Python's from foo import *; we can't use * because that's an infix operator, and the syntax doesn't and won't special-case individual commands (remember that one of the goals of the syntax is not to have hardcoded keywords). Maybe from foo import all(), and also things like from foo import except(foo, bar). I don't know.

Why foo/bar instead of foo.bar? Because I thought that import foo.bar might give the impression that the bindings are to be accessed as foo.bar.hello() (like Python) instead of just bar.hello() (as Hel does). And why I wanted this semantics? Because I was unsure what foo would be when you import foo.bar: a module containing just bar? What if I import foo later? What if foo contains a binding bar itself? Does the module foo have to exist for me to be able to import foo.bar? To avoid all these questions ("do the simple thing"), I decided it would be simpler to import the module without having to deal with the whole hierarchy, and make it available as just bar; and I thought the syntax with / suggested that better.

_____

1 "When in doubt, do the simple thing" has been a sort of mantra in this project. This has helped me avoid analysis paralysis and keep making progress, even though I know eventually I will have to go back and change/improve things. (Note though that the mantra is "do the simple thing", not "do the simplest thing".)

Comentários / Comments

Loops and blocks in Hel

2019-04-11 17:11 -0300. Tags: comp, prog, pldesign, hel, fenius, in-english

Hel got a preliminary version of its main looping and non-local control flow primitives today: do, redo and return. They have characteristics similar to Common Lisp's block/return-from, Scheme's named let, and Clojure's loop/recur (and, one might say, Java's labeled blocks, continues and breaks), though they are not exactly the same as any of these. In this post, I will describe how they work.

do

do creates a labeled block with parameters and initial values. It has the syntax:

do name(var1=value1, var2=value2, ...) {
    body
}

What this does is to evaluate body in an environment containing the specified variables with the given values. The variable declaration section is like a function parameter declaration where all parameters are given default values. For example:

do block(x=1, y=2) {
    x+y
}

will return 3.

Within the block, name is bound to a tag for the block. This tag can be used with the redo and return commands.

redo

redo can be used to repeat the named block, with new values for the block variables. For example:

# Print all integers from 1 to `limit`.
let count_until(limit) = {
    do block(i=1) {             # First iteration will run with `i` = 1.
        if i <= limit {         # If we have not reached the limit yet...
            print(i)            # Print the current value of `i`...
            redo block(i=i+1)   # And repeat the block, with a new value of `i`.
        }
    }
}

This is analogous to Clojure's recur, except it does not have to be in tail position, and you can specify the label of the block you want to repeat (so you can have nested blocks and escape to outermost ones).

This is also similar to Scheme's named let, except that the new execution of the block replaces the current one, rather than behaving like a regular function call.

The names of the parameters in the redo call are optional; we could have written redo block(i+1) instead of redo block(i=i+1). This is analogous to the function call syntax.

return

return can be used to return a value immediately from a block. For example, suppose we have a foreach function which takes a list and a function, and applies the function to each element of the list in order:

let foreach(list, f) = {
    do block(list=list) {                # Start iterating with the full list
        if list != [] {                  # If the list is not empty yet...
            f(list.first)                # Apply the function to the first element...
            redo block(list=list.rest)   # And repeat the block for the remaining ones.
        }
    }
}

Now we want to write a function to test if a given element is in a list. We want to reuse foreach to do the iteration, but we want to stop the iteration (and get out of foreach immediately) when we find the element in the list. We can do this with return:

let is_member(searched, list) = {
    do out() {
        # Call `foreach` with an anonymous function, which will be called
        # for each element of the list.
        foreach(list, fn (item) {
            if item == searched {        # If the current element is the one we are searching...
                return true from out     # Return `true` immediately from the `out` block.
            }
        })
        # If we got here, it's because the element was not found.
        return false from out
    }
}

These constructs can be compared to Java's labeled blocks, continues and breaks. However, Hel blocks take parameters, which must be specified when redoing them (the equivalent of continue); and Hel blocks return a value, which must be specified when returning from them (the equivalent of break).

To do / open questions

Common Lisp has the notion of a default block (which is the block labelled nil). Some constructs, like return, return from the default block, so you can avoid naming the block if you are only using one. It would be nice to have something similar in Hel.

Currently the parameter/argument binding logic for blocks is the same one used for functions. This means that if one of the block's arguments is omitted from the redo call, it will acquire the initial value specified in the beginning of the block! This is most likely not what you want. Alternative behaviours would be to forbid omitting block arguments, or reusing the value in the current iteration rather than the initial value.

Perhaps instead of using special forms for redo and return, these could be methods of the tag object, so we would write, for example, block.redo(i=i+1) instead of redo block(i=i+1), and block.return(42) rather than return 42 from block. I like the special form better, especially for redo because the block name stays together with the arguments just like in the block declaration. It also allows the possibility of omitting the block name if we get default blocks in the future.

Comentários / Comments

Main menu

Posts recentes

Comentários recentes

Tags

em-portugues (213) comp (135) prog (67) in-english (48) life (47) pldesign (34) unix (33) lang (32) random (28) about (27) mind (25) lisp (23) mundane (22) fenius (19) web (17) ramble (17) img (13) rant (12) hel (12) scheme (10) privacy (10) freedom (8) academia (7) esperanto (7) copyright (7) music (7) bash (7) lash (7) home (6) mestrado (6) shell (6) conlang (5) misc (5) emacs (4) politics (4) book (4) php (4) worldly (4) editor (4) latex (4) etymology (4) android (4) film (3) kbd (3) security (3) network (3) wrong (3) c (3) tour-de-scheme (3) poem (2) philosophy (2) comic (2) llvm (2) cook (2) treta (2) lows (2) physics (2) perl (1) german (1) wm (1) en-esperanto (1) audio (1) translation (1) old-chinese (1) kindle (1) pointless (1)

Elsewhere

Quod vide


Copyright © 2010-2020 Vítor De Araújo
O conteúdo deste blog, a menos que de outra forma especificado, pode ser utilizado segundo os termos da licença Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International.

Powered by Blognir.