Why Static Languages Suffer From Complexity

Hirrolot’s Blog

Jan 19, 2022

People in the programming language design community strive to make their languages more expressive, with a strong type system, mainly to increase ergonomics by avoiding code duplication in final software; however, the more expressive their languages become, the more abruptly duplication penetrates the language itself.

This is what I call statics-dynamics biformity: whenever you introduce a new linguistic abstraction to your language, it may reside either on the statics level, on the dynamics level, or on the both levels. In the first two cases, where the abstraction is located only on one particular level, you introduce inconsistency to your language; in the latter case, you inevitably introduce the feature biformity.

For our purposes, the statics level is where all linguistic machinery is being performed at compile-time. Similarly, the dynamics level is where code is being executed at run-time. Thence the typical control flow operators, such as if/while/for/return, data structures, and procedures, are dynamic, whereas static type system features and syntactical macros, are static. In essence, the majority of static linguistic abstractions have their correspondence in the dynamic space and vice versa:

Dynamics Statics
Array Record type/Tuple/Heterogeneous list
Tree (data structure) Sum type/Coproduct
Value Generic/Associated type1
Loop/recursion Type-level induction
If-then-else/Pattern matching Multiple trait implementations
Function signature trait F<In...> { type Out; }
Function implementation impl F<...> for T { type Out = ...; }
Function call <T as F<...>>::Out

In the following sections, before elaborating on the problem further, let me demonstrate to you how to implement logically equivalent programs using both the static and dynamic approaches. Most of the examples are written in Rust but can be applied to any other general-purpose programming language with enough expressive type system; keep in mind that this writeup is language-agnostic and is concerned with the philosophy of PLT in general rather than with particular PL implementations. If you feel busy, feel free to jump right to the main section about the problem explanation.

Record type – Array

Consider your everyday manipulation with record types (playground):

automobile-static.rs

struct Automobile {
    wheels: u8,
    seats: u8,
    manufacturer: String,
}

fn main() {
    let my_car = Automobile {
        wheels: 4,
        seats: 4,
        manufacturer: String::from("X"),
    };

    println!(
        "My car has {} wheels and {} seats, and it was made by {}.",
        my_car.wheels, my_car.seats, my_car.manufacturer
    );
}

The same can be done using arrays (playground):

automobile-dynamic.rs

use std::any::Any;

#[repr(usize)]
enum MyCar {
    Wheels,
    Seats,
    Manufacturer,
}

fn main() {
    let my_car: [Box<dyn Any>; 3] = [Box::new(4), Box::new(4), Box::new("X")];

    println!(
        "My car has {} wheels and {} seats, and it was made by {}.",
        my_car[MyCar::Wheels as usize]
            .downcast_ref::<i32>()
            .unwrap(),
        my_car[MyCar::Seats as usize].downcast_ref::<i32>().unwrap(),
        my_car[MyCar::Manufacturer as usize]
            .downcast_ref::<&'static str>()
            .unwrap()
    );
}

Yes, if we specify an incorrect type to .downcast_ref, we will get a panic. But the very logic of the program remains the same, only we elevate type checking to run-time.

Going further, we can encode static Automobile as a heterogenous list:

automobile-hlist.rs

use frunk::{hlist, HList};

struct Wheels(u8);
struct Seats(u8);
struct Manufacturer(String);
type Automobile = HList![Wheels, Seats, Manufacturer];

fn main() {
    let my_car: Automobile = hlist![Wheels(4), Seats(4), Manufacturer(String::from("X"))];

    println!(
        "My car has {} wheels and {} seats, and it was made by {}.",
        my_car.get::<Wheels, _>().0,
        my_car.get::<Seats, _>().0,
        my_car.get::<Manufacturer, _>().0
    );
}

This version enforces exactly the same type checks as automobile-static.rs, but additionally provides methods for manipulating with type Automobile as with ordinary collections! E.g., we may want to reverse our automobile:2

assert_eq!(
    my_car.into_reverse(),
    hlist![Manufacturer(String::from("X")), Seats(4), Wheels(4)]
);

Or we may want to zip our car with their car:

let their_car = hlist![Wheels(6), Seats(4), Manufacturer(String::from("Y"))];

assert_eq!(
    my_car.zip(their_car),
    hlist![
        (Wheels(4), Wheels(6)),
        (Seats(4), Seats(4)),
        (Manufacturer(String::from("X")), Manufacturer(String::from("Y")))
    ]
);

… And so forth.

However, sometimes we may want to apply type-level computation to ordinary structs and enums, but we cannot do it because we are unable to extract the very structure of a type definition (fields and types/variants and their function signatures) from a corresponding type name, especially if this type is external to our crate and we cannot put a derive macro onto it.3 To resolve the issue, Frunk developers decided to create such a procedural macro that examines the internal structure of a type definition by implementing the Generic trait for it; it has the type Repr associated type, which, when implemented, equals to some form of a manipulatable heterogenous list. Still, all other types (well, transparent ones, such as DTOs) that do not have this derive macro, are left unexaminable, owing to the aforementioned limitations of Rust.

Sum type – Tree

One may find sum types good to represent an AST node (playground):

ast-static.rs

use std::ops::Deref;

enum Expr {
    Const(i32),
    Add(Box<Expr>, Box<Expr>),
    Sub(Box<Expr>, Box<Expr>),
    Mul(Box<Expr>, Box<Expr>),
    Div(Box<Expr>, Box<Expr>),
}

use Expr::*;

fn eval(expr: &Box<Expr>) -> i32 {
    match expr.deref() {
        Const(x) => *x,
        Add(lhs, rhs) => eval(&lhs) + eval(&rhs),
        Sub(lhs, rhs) => eval(&lhs) - eval(&rhs),
        Mul(lhs, rhs) => eval(&lhs) * eval(&rhs),
        Div(lhs, rhs) => eval(&lhs) / eval(&rhs),
    }
}

fn main() {
    let expr: Expr = Add(
        Const(53).into(),
        Sub(
            Div(Const(155).into(), Const(5).into()).into(),
            Const(113).into(),
        )
        .into(),
    );

    println!("{}", eval(&expr.into()));
}

The same can be done using tagged trees (playground):

ast-dynamic.rs

use std::any::Any;

struct Tree {
    tag: i32,
    value: Box<dyn Any>,
    nodes: Vec<Box<Tree>>,
}

const AST_TAG_CONST: i32 = 0;
const AST_TAG_ADD: i32 = 1;
const AST_TAG_SUB: i32 = 2;
const AST_TAG_MUL: i32 = 3;
const AST_TAG_DIV: i32 = 4;

fn eval(expr: &Tree) -> i32 {
    let lhs = expr.nodes.get(0);
    let rhs = expr.nodes.get(1);

    match expr.tag {
        AST_TAG_CONST => *expr.value.downcast_ref::<i32>().unwrap(),
        AST_TAG_ADD => eval(&lhs.unwrap()) + eval(&rhs.unwrap()),
        AST_TAG_SUB => eval(&lhs.unwrap()) - eval(&rhs.unwrap()),
        AST_TAG_MUL => eval(&lhs.unwrap()) * eval(&rhs.unwrap()),
        AST_TAG_DIV => eval(&lhs.unwrap()) / eval(&rhs.unwrap()),
        _ => panic!("Out of range"),
    }
}

fn main() {
    let expr = /* Construction omitted... */;

    println!("{}", eval(&expr));
}

Similarly to how we did with struct Automobile, we can represent enum Expr as frunk::Coproduct. This is left as an exercise to the reader.

Value – Associated type

We may want to negate a boolean value using the standard operator ! (playground):

negate-dynamic.rs

fn main() {
    assert_eq!(!true, false);
    assert_eq!(!false, true);
}

The same can be done through associated types4 (playground):

negate-static.rs

use std::marker::PhantomData;

trait Bool {
    type Value;
}

struct True;
struct False;

impl Bool for True { type Value = True; }
impl Bool for False { type Value = False; }

struct Negate<Cond>(PhantomData<Cond>);

impl Bool for Negate<True> {
    type Value = False;
}

impl Bool for Negate<False> {
    type Value = True;
}

const ThisIsFalse: <Negate<True> as Bool>::Value = False;
const ThisIsTrue: <Negate<False> as Bool>::Value = True;

In fact, the Turing completeness of Rust’s type system is built upon this principle combined with type induction (which we shall see a bit later). Every time you see an ordinary value in terms of Rust, know that it has its formal correspondence on the type-level, in the computational sense. Every time you write some algorithm, it has its correspondence on the type-level, using conceptually equivalent constructions! If you are interested in how, the above article provides a mathematical proof: first, the author implements a so-called language Smallfuck using dynamics: a sum type, pattern matching, recursion, and then using statics: logic on traits, associated types, etc.

Recursion – Type-level induction

Let me show you one more example. But hold on tight this time (playground)!

peano-dynamic.rs

use std::ops::Deref;

#[derive(Clone, Debug, PartialEq)]
enum Nat {
    Z,
    S(Box<Nat>),
}

fn add(lhs: &Box<Nat>, rhs: &Box<Nat>) -> Nat {
    match lhs.deref() {
        Nat::Z => rhs.deref().clone(), // I
        Nat::S(next) => Nat::S(Box::new(add(next, rhs))), // II
    }
}

fn main() {
    let one = Nat::S(Nat::Z.into());
    let two = Nat::S(one.clone().into());
    let three = Nat::S(two.clone().into());

    assert_eq!(add(&one.into(), &two.into()), three);
}

This is the Peano encoding of a natural number. In the add function, we use recursion to compute a sum and pattern matching to find out where to stop.

As recursion corresponds to type induction and pattern matching corresponds to multiple implementations, the same can be done at compile-time (playground):

peano-static.rs

use std::marker::PhantomData;

struct Z;
struct S<Next>(PhantomData<Next>);

trait Add<Rhs> {
    type Result;
}

// I
impl<Rhs> Add<Rhs> for Z {
    type Result = Rhs;
}

// II
impl<Lhs: Add<Rhs>, Rhs> Add<Rhs> for S<Lhs> {
    type Result = S<<Lhs as Add<Rhs>>::Result>;
}

type One = S<Z>;
type Two = S<One>;
type Three = S<Two>;

const THREE: <One as Add<Two>>::Result = S(PhantomData);

Here, impl ... for Z is the base case (termination case), and impl ... for S<Lhs> is the induction step (recursion case) – similar to how we did with pattern matching using match. Likewise, as in the first example, the induction works by reducing the first argument to Z: <Lhs as Add<Rhs>>::Result works just like add(next, rhs) – it invokes pattern matching once again to drive the computation further. Note that the two trait implementations indeed belong to the same logical function implementation; they look detached because we perform pattern matching on our type-level number (Z and S<Next>). This is somewhat similar to what we are used to see in Haskell, where each pattern matching case looks like a separate function definition:

peano-static.hs

import Control.Exception

data Nat = Z | S Nat deriving Eq

add :: Nat -> Nat -> Nat
add Z rhs = rhs -- I
add (S next) rhs = S(add next rhs) -- II

one = S Z
two = S one
three = S two

main :: IO ()
main = assert ((add one two) == three) $ pure ()

Type-level logic reified

The purpose of this writeup is only to convey the intuition behind the statics-dynamics biformity and not to provide a formal proof – for the latter, please refer to an awesome library called type-operators (by the same person who implemented Smallfuck on types). In essence, it is an algorithmic macro eDSL that boils down to type-level manipulation with traits: you can define algebraic data types and perform data manipulations on them similar to how you normally do in Rust, but in the end, the whole code will dwell on the type-level. For more details, see the translation rules and an excellent guide by the same author. Another noteworthy project is Fortraith, which is a “compile-time compiler that compiles Forth to compile-time trait expressions”:

forth!(
    : factorial (n -- n) 1 swap fact0 ;
    : fact0 (n n -- n) dup 1 = if drop else dup rot * swap pred fact0 then ;
    5 factorial .
);

The above code translates a simple factorial implementation to computation on traits and associated types. Later, you obtain a result as follows:

println!(
    "{}",
    <<<Empty as five>::Result as factorial>::Result as top>::Result::eval()
);

Having considered everything above, it is crystal clear that the logic part remains the same, no matter how you call it: be it statics or dynamics.

The unfortunate consequences of being static

Are you quite sure that all those bells and whistles, all those wonderful facilities of your so called powerful programming languages, belong to the solution set rather than the problem set?

Edsger Dijkstra (Edsger Dijkstra, n.d.)

Programming languages nowadays do not focus on the logic. They focus on the mechanisms inferior to logic; they call boolean negation the most simple operator that must exist from the very beginning but negative trait bounds are considered a debatable concept with “a lot of issues.” The majority of mainstream PLs support the tree data structure in their standard libraries, but sum types stay unimplemented for decades. I cannot imagine a single language without the if operator, but only a few PLs accommodate full-fledged trait bounds, not to mention pattern matching. This is inconsistency – it compels software enginners design low-quality APIs that either go dynamic and expose a very few compile-time checks or go static and try to circumvent the fundamental limitations of a host language, thereby making their usage more and more abstruse. Combining statics and dynamics in a single working solution is also complicated since you cannot invoke dynamics in a static context. In terms of function colours, dynamics is coloured red, whereas statics is blue.

In addition to this inconsistency, we have the feature biformity. In such languages as C++, Haskell, and Rust, this biformity amounts to the most perverse forms; you can think of any so-called “expressive” programming language as of two or more smaller languages put together: C++ the language and C++ templates/macros, Rust the language and type-level Rust + declarative macros, etc. With this approach, each time you write something at a meta-level, you cannot reuse it in the host language and vice versa, thus violating the DRY principle (as we shall see in a minute). Additionally, biformity increases the learning curve, hardens language evolution, and finally ends up in such a feature bloat that only the initiated can figure out what is happening in the code. Take a look at any production code in Haskell and you will immediately see those numerous GHC #LANGUAGE clauses, each of which signifies a separate language extension:

feature-bloat.hs

{-# LANGUAGE BangPatterns               #-}
{-# LANGUAGE CPP                        #-}
{-# LANGUAGE ConstraintKinds            #-}
{-# LANGUAGE DefaultSignatures          #-}
{-# LANGUAGE DeriveAnyClass             #-}
{-# LANGUAGE DeriveGeneric              #-}
{-# LANGUAGE DerivingStrategies         #-}
{-# LANGUAGE FlexibleContexts           #-}
{-# LANGUAGE FlexibleInstances          #-}
{-# LANGUAGE GADTs                      #-}
{-# LANGUAGE GeneralizedNewtypeDeriving #-}
{-# LANGUAGE NamedFieldPuns             #-}
{-# LANGUAGE OverloadedStrings          #-}
{-# LANGUAGE PolyKinds                  #-}
{-# LANGUAGE RecordWildCards            #-}
{-# LANGUAGE ScopedTypeVariables        #-}
{-# LANGUAGE TypeFamilies               #-}
{-# LANGUAGE UndecidableInstances       #-}
{-# LANGUAGE ViewPatterns               #-}

Adapted from haskell/haskell-language-server.

When a host language does not provide enough static capabilities needed for convenient development, some programmers go especially insane and create whole new compile-time metalanguages and eDSLs atop of existing ones. Thus, inconsistency has the treacherous property of transforming into biformity:

Sometimes, software engineers find their languages too primitive to express their ideas even in dynamic code. But they do not give up:

Recalling the famous Greenspun’s tenth rule, such handmade metalanguages are typically “ad-hoc, informally-specified, bug-ridden, and slow,” with quite vague semantics and awful documentation. The concept of a metalinguistic abstraction simply does not work, albeit the rationale of creating highly declarative, small domain-specific languages sounds so cool at first sight. When a problem domain (or some intermediate machinery) is expressed in terms of a host language, you need to understand how to chain calls together to get things done – this is what we usually call an API; however, when this API is written in another language, then, in addition to the calling sequence, you need to understand the syntax and semantics of that language, which is very unfortunate for two reasons: the mental burden it lays upon developers and a very limited number of developers that can support such metalanguages. From my experience, handmade metalinguistics tend to quickly go out of hand and spread across the whole codebase, thereby making it harder to dig into. Not only reasoning is impaired but also compiler-developer interaction: have you ever tried to use a sophisticated type or macro API? If yes, then you should be perfectly acquainted with inscrutable compiler diagnostics, which can be summarised in the following screenshot:8

This is woefully to say, but it seems that an “expressive” PL nowadays means “Hey there, I have seriously messed up with the number of features, but that is fine!”

Finally, a word has to be said about metaprogramming in a host language. With such templating systems as Template Haskell and Rust’s procedural macros, we can manipulate an AST9 of a host language using the same language, which is good in terms of biformity but unpleasant in terms of general language inconsistency. Macros are not functions: we cannot partially apply a macro and obtain a partially applied function (or vice versa), since they are just different concepts – this can turn out to be a pain in the ass if we are to design a generic and easy-to-use library API. Personally, I do think that procedural macros in Rust are a giant design mistake that is comparable to #define macros in plain C: aside from pure syntax, the macro system simply has no idea about the language being manipulated; instead of a tool to extend and work with a language gracefully, you get slightly enhanced text substitution and nothing more. E.g., imagine there is an enumeration called Either, whose definition is as follows:

either.rs

pub enum Either<L, R> {
    Left(L),
    Right(R),
}

Adapted from either::Either.

Now imagine we have an arbitrary trait Foo, and we are willing to implement this trait for Either<L, R>, where L and R both implement Foo. It turns out that we cannot apply a derive macro to Either that implements this trait, even if the name is known because, in order to do this, this macro must know all the signatures of Foo. To make the situation even worse, Foo may be defined in a separate library, meaning that we cannot augment its definition with extra meta-information needed for the derivation for Either<L, R>. While it may seem as a rare scanario, in fact it is not; I highly encourage you to look at tokio-util’s Either, which is exactly the same enumeration but it implements Tokio-specific traits, such as AsyncRead, AsyncWrite, AsyncSeek, etc.10 Now imagine you have five different Eithers in your project that came from different libraries – that would be a true integration headache! While type introspection may be a compromise, it would nonetheless make the language even more complex than it already is.

Idris: The way out?

One of the most fundamental features of Idris is that types and expressions are part of the same language – you use the same syntax for both.

Edwin Brady, the author of Idris (Edwin Brady, n.d.)

Let us think a little bit about how to workaround the issue. If we make our languages fully dynamic, we will win biformity and inconsistency,11 but will imminently lose the pleasure of compile-time validation and will end up debugging our programs at mid-nights. The misery of dynamic type systems is widely known.

The only way to approach the problem is to make a language whose features are both static and dynamic and not to split the same feature into two parts. Thus, the ideal linguistic abstraction is both static and dynamic; however, it is still a single concept and not two logically similar concepts but with different interfaces.12 A perfect example is CTFE, colloquially known as constexpr: same code can be executed at compile-time under a static context and at run-time under a dynamic context (e.g., when requesting a user input from stdin.); thus, we do not have to write different code for compile-time (statics) and run-time (dynamics), instead we use the same representation.

One possible solution I have seen is dependent types. With dependent types, we can parameterise types not only with other types but with values, too. In a dependently typed language Idris, there is a type called Type – it stands for the “type of all types,” thereby weakening the dichotomy between type-level and value-level. Having such a powerful thing at our disposal, we can express typed abstractions that are usually either built into a language compiler/environment or done via macros. Perhaps the most common and descriptive example is a type-safe printf that calculates types of its arguments on the fly, so let give us the pleasure of mastering it in Idris!13

First, define an inductive data type Fmt and a way to get it from a format string:

data Fmt = FArg Fmt | FChar Char Fmt | FEnd

toFmt : (fmt : List Char) -> Fmt
toFmt ('*' :: xs) = FArg (toFmt xs)
toFmt (  x :: xs) = FChar x (toFmt xs)
toFmt [] = FEnd

Later, we will use it to generate a type for our printf function. The syntax resembles Haskell a lot and should be comprehensible for a reader.

Now the most interesting part:

PrintfType : (fmt : Fmt) -> Type
PrintfType (FArg fmt) = ({ty : Type} -> Show ty => (obj : ty) -> PrintfType fmt)
PrintfType (FChar _ fmt) = PrintfType fmt
PrintfType FEnd = String

What this function does? It calculates a type based on the input argument fmt. As usual, we case-split fmt into three cases and deal with them separately:

  1. (FArg fmt). Since FArg indicates that we are to provide a printable argument, this case produces a type signature that takes an additional parameter:
    1. {ty : Type} means that Idris will deduce a type ty of this argument automatically (implicit argument).
    2. Show ty is a type constraint that says that ty should implement Show.
    3. (obj : ty) is that printable argument we must provide to printf.
    4. PrintfType fmt is a recursive call that deals with the rest of the input fmt. In Idris, inductive types are managed by inductive functions!
  2. (FChar _ fmt). FChar indicates an ordinary character in a format string, so here we just ignore it and continue with PrintfType fmt.
  3. FEnd. This is the end of input. Since we want our printf to produce a String, we return String as an ordinary type.

Now say we have a format string "*x*", or FArg (FChar ('x' (FArg FEnd))); what type will PrintfType generate? Simple:

  1. FArg: {ty : Type} -> Show ty => (obj : ty) -> PrintfType (FChar ('x' (FArg FEnd)))
  2. FChar: {ty : Type} -> Show ty => (obj : ty) -> PrintfType (FArg FEnd)
  3. FArg: {ty : Type} -> Show ty => (obj : ty) -> {ty : Type} -> Show ty => (obj : ty) -> PrintfType FEnd
  4. FEnd: {ty : Type} -> Show ty => (obj : ty) -> {ty : Type} -> Show ty => (obj : ty) -> String

Cool, now it is time to write the coveted printf:

printf : (fmt : String) -> PrintfType (toFmt $ unpack fmt)
printf fmt = printfAux (toFmt $ unpack fmt) [] where
    printfAux : (fmt : Fmt) -> List Char -> PrintfType fmt
    printfAux (FArg fmt) acc = \obj => printfAux fmt (acc ++ unpack (show obj))
    printfAux (FChar c fmt) acc = printfAux fmt (acc ++ [c])
    printfAux FEnd acc = pack acc

As you can see, PrintfType (toFmt $ unpack fmt) occurs in the type signature, meaning that the whole type of printf depends on the input argument fmt! But what does unpack fmt mean? Since printf takes fmt : String, we should convert it to List Char beforehand because we match this string in toFmt; as far as I know, Idris does not allow matching an ordinary String in the same way. Likewise, we do unpack fmt before calling printfAux, since it also takes List Char as a result accumulator.

Let us examine the printfAux implementation:

  1. (FArg fmt). Here we return a lambda function that takes obj and calls show on it, then appends to acc by the ++ operator.
  2. (FChar c fmt). Just append c to acc and call printfAux once again on fmt.
  3. FEnd. Since acc is typed List Char but we have to return String (according to the last case of PrintfType), we call pack on it.

Finally, test printf:

printf.idr

main : IO ()
main = putStrLn $ printf "Mr. John has * contacts in *." 42 "New York"

This prints Mr. John has 42 contacts in "New York".. But what if we do not provide 42 to printf?

Error: While processing right hand side of main. When unifying:
    ?ty -> PrintfType (toFmt [assert_total (prim__strIndex "Mr. John has * contacts in *." (prim__cast_IntegerInt (natToInteger (length "Mr. John has * contacts in *.")) - 1))])
and:
    String
Mismatch
between: ?ty -> PrintfType (toFmt [assert_total (prim__strIndex "Mr. John has * contacts in *." (prim__cast_IntegerInt (natToInteger (length "Mr. John has * contacts in *.")) - 1))]) and String.

test:21:19--21:68
 17 |     printfAux (FChar c fmt) acc = printfAux fmt (acc ++ [c])
 18 |     printfAux FEnd acc = pack acc
 19 | 
 20 | main : IO ()
 21 | main = putStrLn $ printf "Mr. John has * contacts in *." "New York"
                        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

Warning: compiling hole Main.main

Yeah, Idris detects the error and produces a type mismatch! This is basically how you implement type-safe printf with first-class types. If you are curious about the same thing in Rust, take a look at Will Crichton’s attempt, which relies heavily on heterogenous lists we have seen above. The downside of this approach now should be perfectly clear to you: in Rust, the language of the type system differs from the main language, but in Idris, it is indeed the same thing – this is why we can freely define type-level functions as regular functions returning a Type and invoke them later in type signatures. Moreover, because Idris is dependently typed, you can even compute a type based on some run-time argument, which is impossible in such languages as Zig.

I already anticipate the question: what is the problem of implementing printf with macros? After all, println! works just fine in Rust. The problem is macros. Think for yourself: why a programming language needs heavy-duty macros? Because we may want to extend it. Why may we want to extend it? Because a programming language does not fit our needs: we cannot express something using regular linguistic abstractions, and this is why we decide to extend the language with ad-hoc meta-abstractions. In the main section, I provided an argumentation why this approach sucks – because a macro system has no clue about a language being manipulated; in fact, procedural macros in Rust is just a fancy name for the M4 preprocessor. You guys integrated M4 into your language. Of course, this is better than external M4, but it is nevertheless a method of the 20’th century; proc. macros even cannot manipulate an abstract syntax tree, because syn::Item, a common structure used to write proc. macros, is indeed known as a concrete syntax tree, or “parse tree.” On the other hand, types are a natural part of a host language, and this is why if we can express a programmatic abstraction using types, we reuse linguistic abstractions instead of resorting to ad-hoc machinery. Ideally, a programming language should have either no macros or only a lightweight form of syntax rewriting rules (like Scheme’s extend-syntax or syntax extensions of Idris), in order to keep the language consistent and well-suited to solve expected tasks.

That being said, Idris erases the first biformity “values-generics” by introducing Type, the “type of all types.” By doing so, it also resolves a bunch of other correspondences, such as recursion vs. type-level induction, functions vs. trait machinery, and so forth; this, in turn, allows to program in the same language as much as possible, even when dealing with highly generic code. E.g., you can even represent a list of types as List Type, just like List Nat or List String, and deal with it as usual! It is possible due to a cumulative hierarchy of universes: simply speaking, Type is of type Type 1, Type 1 is of type Type 2, and so on. Since Data.List’s generic named a is “implicitly” typed Type, it can be Nat and String as well as Type; in the latter case, a would be deduced as Type 1. Such an infinite sequence of types is needed to avoid a variation of Russell’s paradox by making an inhabitant “structurally smaller” than its type.

Idris, however, is not a simple language. Our twenty line example of printf already uses a “whole lotta features,” such as inductive data types, dependent pattern matching, implicits, type constraints, to mention a few. Additionally, Idris features computational effects, elaborator reflection, coinductive data types, and many-many stuff for theorem proving. With such a pleiad of typing facilities, you typically twiddle with your language machinery rather than doing some meaningful work. I can hardly believe that in their current state, dependent types will find massive production use; as for now, in the programming world, they are no more than a fancy thing for PL researchers and random enthusiasts like me. Dependent types alone are just too low-level.

Zig: Simpler, but too systems

In Zig, types are first-class citizens. They can be assigned to variables, passed as parameters to functions, and returned from functions.

The Zig manual (Zig developers, n.d.)

Our last patient would be the Zig programming language. Here is a compile-time printf implementation in Zig (sorry no highlighting yet):

printf.zig

const std = @import("std");

fn printf(comptime fmt: []const u8, args: anytype) anyerror!void {
    const stdout = std.io.getStdOut().writer();

    comptime var arg_idx: usize = 0;

    inline for (fmt) |c| {
        if (c == '*') {
            try printArg(stdout, args[arg_idx]);
            arg_idx += 1;
        } else {
            try stdout.print("{c}", .{c});
        }
    }

    comptime {
        if (args.len != arg_idx) {
            @compileError("Unused arguments");
        }
    }
}

fn printArg(stdout: std.fs.File.Writer, arg: anytype) anyerror!void {
    if (@typeInfo(@TypeOf(arg)) == .Pointer) {
        try stdout.writeAll(arg);
    } else {
        try stdout.print("{any}", .{arg});
    }
}

pub fn main() !void {
    try printf("Mr. John has * contacts in *.\n", .{ 42, "New York" });
}

Here, we use a feature called comptime: a comptime function parameter means that it must be known at the time of compilation. Not only it allows for aggressive optimisations but also opens a Valhalla of “metaprogramming” facilities, most notably without separate macro-level or type-level sublanguages. The above code needs no further explanation – the mundane logic should be clear to every programmer, unlike printf.idr that seems to look like a fruit of a mad genius’ wet fantasies.

If we omit 42, Zig will report a compilation error:

An error occurred:
/tmp/playground2454631537/play.zig:10:38: error: field index 1 outside tuple 'struct:33:52' which has 1 fields
            try printArg(stdout, args[arg_idx]);
                                     ^
/tmp/playground2454631537/play.zig:33:15: note: called from here
    try printf("Mr. John has * contacts in *.\n", .{ "New York" });
              ^
/tmp/playground2454631537/play.zig:32:21: note: called from here
pub fn main() !void {
                    ^

The only inconvenience I experienced during the development of printf is massive errors… Much like C++ templates. However, I admit that this can be solved (or at least workarounded) by more explicit type constraints. Overall, the design of Zig’s type system seems reasonable: there is a type of all types called type, and using comptime, we can compute types at compile-time via regular variables, loops, procedures, etc. We can even perform type reflection through the @typeInfo, @typeName, and @TypeOf built-ins! Yes, we can no longer depend on run-time values, but if you do not need a theorem prover, probably full-blown dependent types are a bit of overkill.

Everything is good except that Zig is a systems language. On their official website, Zig is described as a “general-purpose programming language,” but I can hardly agree with this statement. Yes, you can write virtually any software in Zig, but should you? My experience in maintaining high-level code in Rust and C99 says NO. The first reason is safety: if you make a systems language safe, you will make programmers deal with borrow checker and ownership (or equivalent) issues that have absolutely nothing to do with business logic (believe me, I know the pain); otherwise, if you choose the C-way manual memory management, you will make programmers debugging their code for long hours with the hope that -fsanitize=address would show something meaningful. Moreover, if you want to build new abstractions atop of pointers, you will end up with &str, AsRef<str>, Borrow<str>, Box<str>, and the similar. Come on, I just want a UTF-8 string; most of the time, I do not really care whether it is one of those alternatives.

The second reason is concerned with a language runtime:14 for a language to be systems, to avoid hidden performance penalties, it should have a minimum runtime – no default GC, no default event loop, etc., but for particular applications, it might be necessary to have a runtime – for asynchronous ones, for instance, so actually you must deal with custom runtime code in some way. Here we encounter a whole new set of problems regarding function colours: e.g., having async in your language and having no tools to abstract over synchronous and asynchronous functions means that you divided your language into two parts: synchronous and asynchronous, and say, if you have a generic higher-order library, it will be inevitably marked async to accept all kinds of user callbacks. To resolve the issue, you need to implement some form of effect polymorphism (e.g., monads or algebraic effects), which is still a research topic. High-level languages have innately fewer problems to deal with, and this is why most of the software is written in Java, C#, Python, and JavaScript; in Golang, conceptually, every function is async, thus facilitating consistency by default, without resorting to sophisticated type features. On the contrary, Rust is already considered a complex language and still has no standard means to write truly generic asynchronous code.

Zig can still be used in large systems projects like web browsers, interpreters, and operating system kernels – nobody wants these things to freeze unexpectedly. Zig’s low-level programming features would facilitate convenient operation with memory and hardware devices, while its sane approach to metaprogramming (in the right hands) would cultivate understandable code structure. Bringing it to high-level code would just increase the mental burden without considerable benefits.

Progress is possible only if we train ourselves to think about programs without thinking of them as pieces of executable code.

Edsger Dijkstra

Final words

Static languages enforce compile-time checks; this is good. But they suffer from feature biformity and inconsistency – this is bad. Dynamic languages, on the other hand, suffer from these drawbacks to a lesser extent, but they lack compile-time checks. A hypothetical solution should take the best from the both worlds.

Programming languages ought to be rethought.

References

Edsger Dijkstra. n.d. “On the Teaching of Programming, i.e. On the Teaching of Thinking.” https://www.cs.utexas.edu/users/EWD/transcriptions/EWD04xx/EWD473.html.
Edwin Brady. n.d. “Type-Driven Development with Idris.” Manning Publications Co. https://www.manning.com/books/type-driven-development-with-idris.
Zig developers. n.d. “Introducing the Compile-Time Concept.” https://ziglang.org/documentation/0.9.0/#Introducing-the-Compile-Time-Concept.

  1. In such systems as Calculus of Constructions, polymorphic functions accept generics as ordinary function parameters of the kind *, or “the kind of all types.”↩︎

  2. Need to augment the field types with #[derive(Debug, PartialEq)].↩︎

  3. In addition to inextensibility, derive macros possess tight coupling: if DeriveX is put onto Foo, then foo.rs inevitably depends on DeriveX. This can complicate migration of components, slow down compilation time, and cause merge conflicts.↩︎

  4. We could even generalise these two implementations of Negate over a generic value Cond, but this is impossible due to a known bug in the Rust’s type system.↩︎

  5. In the C++ community, there is an analogous library Boost/Preprocessor.↩︎

  6. Do you remember our variables-generics correspondence?↩︎

  7. Some time ago, minimal working const generics were stabilised. In perspective, they could replace Typenum by using the same integer representation as in ordinary code.↩︎

  8. Got it while working on teloxide, IIRC.↩︎

  9. Technically speaking, Rust cannot work with its true AST, but we will come back to it later.↩︎

  10. It is even more of comedy that initially, I wrote a third-party crate called tokio-either, which just contained that Either with several trait implementations. Only later, the Tokio maintainers decided to move it to tokio-util.↩︎

  11. Terra is a perfect example of a simple dynamic language. In the “Simplicity” section, they show how features of static PLs can be implemented as libraries in dynamic languages.↩︎

  12. Multiple personality disorder? 🤨↩︎

  13. The full code can be found at my gist. I use Idris2, which you can download here.↩︎

  14. For our purposes, a language runtime is some hidden machinery responsible for mapping the semantics of a language (i.e., its abstract machine) to the semantics of a real executor.↩︎