While working with a dynamically typed lang, I came across this:
hash(password, algorithm, algorithmOptions)
Where algorithm
is a constant signaling which hashing algorithm to use, and algorithmOptions
is a dict whose keys depend on algorithm
.
So I thought, can we dictate that if a previous parameter has this value, then this parameter has to have this other value?
E.g.
enum HashAlgo {
Bcrypt,
Argon2,
}
type BcryptOptions = {
Int optionA,
Int optionB,
}
type Argon2Options = {
String optionC,
String optionD,
}
// Here I make this type "depend" on an argument of type HashAlgo
type HashOptions = [HashAlgo] => {
HashAlgo::Bcrypt => BcryptOptions,
HashAlgo::Argon2 => Argon2Options,
}
fun hash(
String password,
HashAlgo algorithm,
// Here I use HashOptions, passing the prev. argument
HashOptions[algorithm] options,
)
This way the compiler can ensure the correct dict is used, based on the value of algorithm
Does something like this exist? I now realize that it would be impossible to type check in compile time based on a runtime value, but if it was allowed only for constants? What do you think?
Multiple ways you can do this. Most of these should also extend to multiple arguments, and although the constant is promoted to type level, you can pass it around nested functions as a type parameter.
With generics
In Java (personally I think this approach is best way to implement your specific example; also Kotlin, C#, and some others are similar):
interface HashAlgo<Options> { String hash(String password, Options options); } class Bcrypt implements HashAlgo<BcryptOptions> { ... } class Argon2 implements HashAlgo<Argon2Options> { ... } record BcryptOptions { ... } record Argon2Options { ... }
In Haskell without GADTs (also Rust is similar):
class HashAlgo opts where hash :: String -> opts -> String data BcryptOptions = BcryptOptions { ... } data Argon2Options = Argon2Options { ... } instance HashAlgo BcryptOptions where hash password BcryptOptions { .. } = ... instance HashAlgo Argon2Options where hash password Argon2Options { .. } = ...
In C (with
_Generic
):typedef struct { ... } bcrypt_options; typedef struct { ... } argon2_options; char* hash_bcrypt(const char* password, bcrypt_options options) { ... } char* hash_argon2(const char* password, argon2_options options) { ... } #define hash(password, options) _Generic((options), bcrypt_options: hash_bcrypt, argon2_options: hash_argon2)(password, options)
In TypeScript, inverting which type is parameterized (see this StackOverflow question for another TypeScript approach):
type HashAlgo = 'bcrypt' | 'argon2' type HashOptions<H> = H extends 'bcrypt' ? BcryptOptions : H extends 'argon2' ? ArgonOptions : never function hash<H>(password: string, algorithm: H, options: HashOptions<H>): string { ... }
With constant generics or full dependent types
This way is a bit more straightforward but also way more complicated for the compiler, and most languages don’t have these features or they’re very experimental. Dependent types are useful when your constant is non-trivial to compute and you can’t even compute it fully, like vectors with their length as a type parameter and
append
guarantees the return vector’s length is the sum. In that case generics aren’t enough. Constant generics aren’t full dependent types but let you do things like the vector-sum example.In Haskell with GADTs AKA Generic Algebraic Data types (also works in Idris, Agda, and other Haskell-likes; you can simulate in Rust using GATs AKA Generic Associated Types, but it’s much uglier):
data BcryptOptions = BcryptOptions { ... } data Argon2Options = Argon2Options { ... } data Algorithm opts where Bcrypt :: Algorithm BcryptOptions Argon2 :: Algorithm Argon2Options hash :: String -> Algorithm opts -> opts -> String hash password algo opts = case algo of | Bcrypt -> ... -- opts is inferred to be BcryptOptions here | Argon2 -> ... -- opts is inferred to be Argon2Options here
In Coq (also flipping the parameterized types again):
Inductive algorithm : Set := bcrypt | argon2. Inductive algorithm_options (A: algorithm) : Set := bcrypt_options : ... -> algorithm_options bcrypt | argon2_options : ... -> algorithm_options argon2. Fixpoint hash (password : string) (algo : algorithm) (opts : algorithm_options also) : string := ... .
This is possible with c++ templates.
I think what you are referring to is “dependent types” They usually occur in more advanced functional programming languages like Idris or Agda but Zig also has them in a way
This would be trivial in python with something like
from typing import overload from enum import Enum @overload def hash(password: str, algorithm: BCryptAlgorithm, options: BCryptOptions): ... @overload def hash(password: str, algorithm: Argon2Algorithm, options: Argon2Options): ... def hash(password: str, algorithm, options): [...implementation...]
Of course it’s python, so at runtime it wouldn’t matter, but the static type checker would complain if you called
hash
withBCryptAlgorithm
andArgon2Options
. You could also have it return different types based on the arguments and then in call sites it’d know which type will be returned based on the type of the arguments. And only the last function has am implementation, the@overload
ones are just type signatures.It’s documented here.