python:
a: str = 1
python:
a: str = 1
Because sometimes that let can be replaced by other things like const. Which can be managed statically by the machine and not by my (imperfect) ability to know if it's mutated or not
I think you can do const thing = ... as constto lock down the mutation?
So is let in some languages. In Rust, you have to constantly opt out from immutability with let mut, which makes writing more procedural code feel like you're fighting with the compiler, and otherwise I don't really see the rationale behind full functional coding. I only had a bug caused only once by unwanted mutation, the hardest part fixing it was to learn the proper use of my debugger tool.
Ok but, in the second example you typically just put final or const in front of the type to denote immutability. I still don't see the advantage to the first declaration.
oh for sure, but I think that's the rarer case for language implementions. Having a consistent structure with alternative keywords in static positions is just easier to develop an AST for. Personally my favorite language doesn't even allow for const values (except by convention) so it's really just a matter of preference
Is it rarer? I think a lot of modern languages go for the first option but pretty much all C style languages use the latter. It's probably a wash for which is more popular I'd think.
Good, now invent a keyword for variables you don't want to declare the type. And now that you have a mix of keywords and identifiers on the same place, you can never update your language again.
Also, make the function declarations not use a keyword too, so you get the full C-style madness of code that changes meaning depending on what libraries you import.
I don't understand how not using a keyword to define a function causes the meaning to change depending on imports. I've never run into an issue like that before. Can you give an example?
Some declarations terminate on the name, other declarations go one requiring more tokens. In C, the only thing that differentiates them is the type.
Parenthesis in particular are completely ambiguous. But asterisks and square brackets also create problems.
I have never heard of this problem for C. Can you elaborate or point to some articles?
The basic problem is that identifiers can be either types or variables, and without a keyword letting you know what kind of statement you're dealing with, there's no way of knowing without a complete identifier table. For example, what does this mean:
foo * bar;
If foo is a type, that is a pointer declaration. But if it's a variable, that's a multiplication expression. Here's another simple one:
foo(bar);
Depending on how foo is defined, that could be a function call or a declaration of a variable bar of type foo, with some meaningless parentheses thrown in.
When you mix things together it gets even more crazy. Check this example from this article:
foo(*bar)();
Is bar a pointer to a function returning foo, or is foo a function that takes a bar and returns a function pointer?
let and fn keywords solve a lot of these ambiguity problems because they let the parser know what kind of statement it's looking at, so it can know whether identifiers in certain positions refer to types or variables. That makes parsing easier to write and helps give nicer error messages.
I feel this is related, and hightlight this even further, look at all the ways to initialize something in C++.
https://www.youtube.com/watch?v=7DTlWPgX6zs
If you are really lazy, have a look at making an int at around 7:20. It's not horrible that alone, but it does show how many meanings each thing has with very little difference, added on top of years of legacy compatability accumulation. Then it further goes into detail about the auto use, and how parantheses, bracket, squiggly bracket all can be used and help with the mess.
In C#, you can use 'var' to have an impilict type variable.
String name = ""
var name = ""
So, a keyword
So I think it's still probably unclear to people why "mix of keywords and identifiers" is bad: it means any new keyword could break backwards compatibility because someone could have already named a type the same thing as that new keyword.
This syntax puts type identifiers in the very prominent position of "generic fresh statement after semicolon or newline"
..though I've spent like 10 minutes thinking about this and now it's again not making sense to me. Isn't the very common plain "already_existing_variable = 5" also causing the same problem? We'd have to go back to cobol style "SET foo = 5" for everything to actually make it not an issue
any new keyword could break backwards compatibility
Wouldn't that happen anyway with variable and function names? Any type other than primitive/built in ones are usually camel case so lower case keywords are more likely to clash with single word variable and function names, unless you restrict the cases of those too or allow keyword overriding or something.
Yeah, it's in my edit I realized the same thing. I'm thinking it doesn't actually really make sense and the real reason is more "the specific way C does it causes a lot of problems so we're not poking syntax like that with a 10 foot pole" + "it makes writing the parser easier" + maybe a bit of "it makes grepping easier"
One thing that annoyed me about C# as a Java guy is that it really wants you to use camel case for function and property names, even private ones. I don't like it specifically because it's hard to differentiate between a function/property and a type.
But C# has quite a few keywords and seem to like adding them more than Java.
Maybe that's their way of ensuring keywords don't clash with stuff?
Ah I was misunderstanding the problem. And learned something new about C#, seems in order to avoid breaking existing code they introduce "contextual keywords" var being added later, it is a contextual. You can create a class 'var' and the compiler will prefer it.
At least in C#, you can define variables with keyword names like this:
var @struct = "abc"
I think in Kotlin you can do the same, and even include spaces with backticks like val abstract class = "abc"
I'm not sure if other languages allow that, regardless it should be rarely used.
Swift also uses backticks and Rust has a dumb one in the form of r#thekeyword. Still much better than introducing a async as a new keyword in a minor version of a language and breaking a bunch of libraries.
Good, now invent a keyword for variables you don't want to declare the type.
auto. Also in D, you only need const if you don't want to specify a type for a constant, the compiler automatically inferres it to you.
Function declarations can be easily decyphered from context, no problem.
IT'S SHOWTIME
I NEED YOUR CLOTHES YOUR BOOTS AND YOUR MOTORCYCLE a
GET TO THE CHOPPER a
HERE IS MY INVITATION "ArnoldC is the best."
ENOUGH TALK
TALK TO THE HAND a
YOU HAVE BEEN TERMINATED
let a: &'static str
Rust is verbose, but C++ might still take the cake with its standard library templates. Especially when using fully-qualified type names...
auto a = ::std::make_shared<::std::basic_string<char, ::std::char_traits<char>, MyAllocator<char>>>();
A reference-counted shared pointer to a string of unspecified character encoding and using a non-default memory allocator.
fully qualified type names make any language awful.
Here's the same example in rust:
let a = std::rc::Rc::new(std::vec::Vec<u8, MyAllocator>::new());
I believe u8 also comes from a module, so it would be something like std::u8::u8, but I'm not sure.
No one would ever write that, though. The following is equivalent:
auto a = std::make_shared<std::string>()
std::string doesn't have a template type for the allocator. You are stuck using the verbose basic_string type if you need a special allocator.
But, of course, nobody sane would write that by hand every time. They would use a typedef, like how std::string is just a typedef for std::basic_string<char, std::char_traits<char>, std::allocator<char>>. Regardless, the C++ standard library is insanely verbose when you start dropping down into template types and using features at an intermediate level. SFINAE in older versions of C++ was mindfuck on the best of days, for example.
Don't get me wrong, though. I'm not saying Rust is much better. Its saving grace is its type inference in let expressions. Without it, chaining functional operations on iterators would be an unfathomable hellscape of Collect<Skip<Map<vec::Iter<Item = &'a str>>>>
Yeah, I missed the custom allocator at first. I thought I deleted my comment fast enough, but I guess you were faster. :)
As long as you call .collect() on it at the end, don't need to write the entire type as it is a method with a generic parameter, and returns that generic.
The intermediate iterators though, those are hell. Especially if you pass it to a lambda and for some reason rust can't infer the type.
I once came across a parsing library that did the parsing with basically just the type system. It was absolute hell to debug and build. Types of parsers would be hundreds of characters long. It would take tens of minutes to build a simple parser.
I don't know much much time it would take to build a complex parser, since it was unable to. it reached the max type generic depth of the rust compiler, and would just refuse to compile.
I believe it was called Chomsky or Chumsky or something like that. Discovering nom was a blessing.
Yeah, I mean Rust is only verbose if you want it to be. let foo = "bar"; is valid rust too, no need to declare the type and definitely no need to declare the lifetime.
For that matter, if you ever declare something as explicitly 'static in code that isn't embedded or super optimized, you're probably doing it wrong.
I don't understand why verbose is bad. Verbose is maintainable.
Not when taken to such an extreme so as to obfuscate the meaning and behavior of code, and make it difficult to understand how you would arrive at that code.
Sane defaults serve to reduce verbosity without obfuscating meaning, simpler syntax with different ordering and fewer tokens reduce verbosity to make the code easier to read by reducing the amount of text you have to pay attention to to understand what the result is.
I imagine there's also a distinction to be made between verbosity and redundancy - sometimes extra text might fail to carry information, or carry information that's already carried elsewhere. I'm not sure where the line should be drawn, because sometimes duplicate information can be helpful, and spacing out information with technically meaningless text has value for readability, but I feel like it's there.
Frankly, I have way more often ended up scratching my head for some cryptic piece of syntax than the opposite. Sure I could sit down and spend the rest of my life learning each language designer's favourite syntax optimisations, but I kinda don't want to. I'm a human, not a parser.
And frankly, a good IDE editor should be able to fold/unfold verbose syntax. Better for it then to be there and folded, than for it to not be there and for someone to have to work backwards the intendrd meaning of every single line of code.
String a: new String()
"" most likely
"".to_string() probably
I am not married to any particular language's syntax choices but sure if we want to play "make it feel like C++":
std::cpp26::string a : std::cpp26::make_unique( std::cpp26::string{} );
I could try and pass it through a number of iostreams or a to_array conversion to see if I can push a to_string() in somewhere...
I've always wondered where all this 'let' business started
More than you'd ever want to know: https://en.m.wikipedia.org/wiki/Let_expression
It's commonly used in math to declare variables so I assume programming languages borrowed it from there.
BASIC uses (used?) it to declare variables. (I don't know if earlier languages did.)
Not that that's a reason for other languages to copy it.
Doesn't Basic use Dim a As String?
Older variants used DIM for arrays and LET for other variables. DIM was originally called that because it was setting the dimensions of the array.
In modern BASIC variants, DIM has become a backronym: "declare in memory".
In modern BASIC variants, DIM has become a backronym: “declare in memory”.
TIL. I always thought it was a backronym for declare in (yo) momma.
Even older variants required both a let to declare the variable and a dim to set its size.
I remember a REDIM command, but I really can't remember what basic it's from.
More specifically, they're borrowing the more mathematical meaning of variables, where if you say x equals 5, you can't later say x is 6, and where a statement like "x = x + 1" is nonsense. Using "let" means you're setting the value once and that's what it's going to remain as long as it exists, while "var" variables can be changed later. Functional languages, which are usually made by very math-y people, will often protest the way programmers use operators by saying that = is strictly for equality and variable assignment is := instead of == and = in most C-style languages.
The actual reason why let ... in syntax tends to not use C-style "type var" like syntax is because it's derived from the syntax type theory uses, and type theorists know about parameterised types. Generics, in C++ parlance, excuse my Haskell:
let foo :: Map Int String = mempty
We have an empty map, and it maps integers to Strings. We call it foo. Compare:
Map Int String foo = mempty
If nothing else, that's just awkward to read and while it may be grammatically unambiguous (a token is a name if it sits directly in front of =) parser error messages are going to suck. Map<Int,String> is also awkward but alas that's what we're stuck with in Rust because they reasoned that it would be cruel to put folks coming from C++ on angle bracket withdrawal. Also Rust has ML ancestry don't get me started on their type syntax.
There is also the thing where the compiler might mistake your c++ style variable declaration for a function, e.g.
String myfunction():
String myvariable();
How do you do nested parameterized types without it becoming ambiguous though? That's IMO the biggest advantage of the bracket syntax. For example: Map<Tuple<Int, Int, Int> Int>
Map (Int, Int) Int. Kind of a bad example because tuples have special-case infix syntax, the general case would be Map Int (Either Int Bool). Follows the same exact syntax as function application just that types (by enforced convention) start upper case. Modulo technical wibbles to ensure that type inference is possible you can consider type constructors to be functions from types to types.
...function application syntax is a story in itself in Haskell because foo a b c gets desugared to (((foo a) b) c): There's only one-argument functions. If you want to have more arguments, accept an argument and return a function that accepts yet another argument. Then hide all that under syntactic sugar so that it looks innocent. And, of course, optimise it away when compiling. Thus you can write stuff like map (+5) xs in Haskell while other languages need the equivalent of map (\x -> x + 5) xs (imagine the \ is a lambda symbol).
Until now, I looked at let and thought, "maybe they just felt like doing it that way".
Makes a lot more sense now.
Not to short-circuit the joke, but in this case, it's because the valid JavaScript version is...
let a
...and one of TypeScript's main design goals is to be a superset of JavaScript, that only adds syntax, and doesn't re-write it.
Beyond that, it's probably a case of some new language just using what the designer is familiar with.
top 79 comments