Swift Generics: More specialized than generic? - xcode

XCode 6: Beta 5:
Goal:
I am trying to write generic code for types that are semantically compatible but do not share (or appear to share) sufficient protocols to base my generics on a subset of shared protocols. So far, I have not been able to find a solution, and am wondering I am missing something or if it is a limitation of the language - any insight is appreciated.
Problem:
I have some functions that differ only by type and not by semantics and seem like a natural fit for generics. The problem that I am having, is that from what I can tell, Swift does what seems like parse-time binding of generics, failing if there could conceivably be a problem, and not when there actually is one.
Example:
Consider the following generic functions in a contrived example:
func defVal<T where T:FloatingPointType, T:FloatLiteralConvertible>(T.Type) -> T {
return 0.0
}
func defVal<T:IntegerLiteralConvertible>(T.Type) -> T {
return 0
}
Note that I have provided functions that should span the cases of integers and floats, and intentionally did not want to provide an exhaustive list of all possible variations that are of no relevance to me.
I then want to define generic code that spans types - in this example, int and float types. Note that this code fails to compile even in the absence of any code that calls it:
func doSomethingGeneric<T>(t:T) -> [T]
{
let a = defVal(T) // Type 'T' does not conform to protocol FloatLiteralConvertible
let b = a * a // works
return [t]
}
In my recollection, this would compile in C++ until you called it with an incompatible type, at which point the compiler would catch it.
I also tried other variants of diminished utility:
func doSomethingWithFloats<T
where T:FloatingPointType, T:FloatLiteralConvertible>(t:T) -> [T]
{
let a = defVal(T) // works
let b = a * a // T is not convertible to UInt8
// - need a floating point arithmetic type?
let c = -a // T is not convertible to Float
let f:Float = -a // T is not convertible to Float
return [t]
}
Given the sense that Swift provides protocols as a way of grouping concrete instances (specialized, not generic), I concocted a protocludge:
protocol Initializable {}
extension Float : Initializable {}
extension Double : Initializable {}
extension CGFloat : Initializable {}
func doSomethingWithInitializable<T:Initializable>(t:T) -> [T]
{
let a = defVal(T) // Type 'T' does not conform to protocol FloatLiteralConvertible
let b = a * a // works
return [t]
}
Note that this fails even though FloatLiteralConvertible is implemented across the set of all Initializable types. Put another way, Swift seems to be binding the generic types too early, and treating generic types as if they were specialized concrete instances instead of a greater pattern that would compile out further down the chain. Furthermore, note that while I could derive from FloatLiteralConvertible, this would preclude me from supporting int types etc. If there was a common ArithmeticType protocol, that could conceivably work, but I do not see anything of the sort. And this is the crux of the problem - there is no common protocol that works for both, even though both ints and floating types are semantically compatible (have the same rules of math).
So in summary, how do you write generic functions for which the types are semantically compatible, but for which there are not enough spanning protocols to filter by protocol (in this case - FloatingPointType does not implement IntegerArithmeticType, but is semantically capable of arithmetic).
Thanks in advance.

Unlike C++, Swift does not deal with generics by substituting the concrete types at the call site and making a non-generic copy (at least in general it doesn't - as an optimization that's allowed, but I digress)
Swift deals with genericity by passing in metadata information describing the actual types at each invocation into one master function, which then uses metadata-provided entry points to manipulate your objects
In your example, Initializable does not provide any operations, so when the compiler tries to execute defVal(T) it has no clue what to do (how can it ensure that there is an overload of defVal for your type?)
What you want to do is actually define defVal as a static function on the Initializable protocol, and then implement it in the extensions, then Swift will know that T.defVal() means something akin to
metadata[Initializable]->defVal(metadata)
Oh, since you're trying to execute a *(T,T), you might also want to make a Multipliable protocol and then your T will be typed as
<T: protocol<Initializable,Multipliable>>

Related

Is F# Constructed Type syntax special?

I was curious about F#'s "constructed type" syntax. It's documented here.
type-argument generic-type-name
or
generic-type-name
With the following examples:
int option
string list
int ref
option<int>
list<string>
ref<int>
Dictionary<int, string>
I was curious if there's anything special about the "backwards" syntax, with the parameter before the type, or if it's just sugar for generic types with one parameter. The following is valid:
type 'a MyOption = // MyOption<'a> also works
| MySome of 'a
| MyNone
But I could not get it to work with multiple type parameters. Why do F# developers prefer this syntax for types with one parameter? Is it possible or desirable to make it work with two?
The backwards syntax is a legacy from OCaml. Personally, I never use it. If you really want to, you can make it work with multiple type arguments like this:
type MyMap = (int, string) Map
However, this generates a pointed warning (that might soon become an error):
This construct is for ML compatibility. The syntax '(typ,...,typ) ident' is not used in F# code. Consider using 'ident<typ,...,typ>' instead. You can disable this warning by using '--mlcompatibility' or '--nowarn:62'.
Bottom line, I would recommend always using .NET syntax instead: MyOption<'a> instead of 'a MyOption.
Why do F# developers prefer this syntax for types with one parameter?
Not all of us do. I love F# and am in awe of it, but find the OCaml style distracting.
It gets especially confusing when the two styles are mixed - compare the readability of Async<Result<int,string list>> list with that of List<Async<Result<int,List<string>>>>.
Here is a thread with some arguments from both sides from fslang suggestions, which I think led to the deprecation of OCaml-style for everything but list, option and a few others.
I find it regrettable that the OCaml style is specified as the preferred option (for these types) in the various style guides, and used throughout the core libraries, while there is such a strong drive to make the language more accessible to newcomers. It definitely adds to the learning curve, as documented in this question,
and here,
here,
here,
here,
here.
Is it possible or desirable to make [OCaml style naming] work with two [type parameters]?
I think a better question is: "Is it possible to only use .NET style?".
Unfortunately the tooling shows types the way they are declared, and the core libraries consistently use OCaml style. I have asked Rider about always showing declarations .NET style in code vision, who referred me to FSharp compiler services. I have not (yet) investigated that avenue further.
In our own code we have taken to overriding the OCaml signatures of functions that ship with F# and other libraries as we come across them, for example:
[<AutoOpen>]
module NoCaml =
module List =
/// Returns a new collection containing only the elements of the collection for which the given predicate returns "true"
let filter = List.filter : ('a -> bool) -> List<'a> -> List<'a>
/// val groupBy : projection:('T -> 'Key) -> list:'T list -> ('Key * 'T list) list (requires equality and equality) 'T is 'a 'Key is 'b Applies a key-generating function to each element of a list and yields a list of unique keys. Each unique key contains a list of all elements that match to this key.
let groupBy = List.groupBy : ('a -> 'b) -> List<'a> -> List<'b * List<'a>>
// etc.
This solves the problem in almost all cases (some exceptions like list construction using [] remain, and need to be overridden at the point of declaration).
I'm not sure what influence this has on performance at runtime - hopefully the extra function calls are optimised away.

Initialize struct with global scope where type depends on boolean

I started using Golang recently and stumbled across a problem:
I have two structs, human and alien, which are both based on the creature struct. I want to initialize one of them based on the value of the isAlien boolean inside of an if-statement.
Using the human := human{} notation or the alien equivalent inside the if blocks to initialize, the instances aren't accessible from outside of the if-statement.
On the other hand, the usual solution of declaring the type and the name of the variable before the if-statement and initializing the variable inside the if-statement doesn't work, because there two are different types:
var h human //use human or alien here?
if isAlien {
h = alien{} //Error: incompatible types
} else {
h = human{}
}
//same when swapping human with alien at the declaration
I know that I could just declare both types before the if-statement but that solution doesn't seem elegant to me.
Is there some obvious solution that I'm missing here?
As you noted, the problem is clearly represented by this statement:
var h human //use human or alien here?
If you plan to use that h variable there after creating the objects, then the type of h must be one that can accept either a human or alien as a value.
The way to do this in Go is by using an ìnterface that both alien and human can fulfil.
So you should declare an interface like:
type subject interface {
// you should list all functions that you plan to use on "h" afterwards
// both "human" and "alien" must implement those functions
}
Then:
var h subject
Will do the trick.
So, I'm going to go out on a limb and say you're probably thinking about this the wrong way.
The first question that occurs to me looking at your example is: what's the return type of this function? In other words, what signature do you need h to be? If alien has an embedded struct creature (which seems to be the inheritance pattern you're trying to follow), and you return a human from your function after declaring h to be a creature, anything that consumes your function will only know that it's dealing with a creature, so there's no point in declaring it a human or an alien in the first place.
I suspect that what you really want to be doing is moving away from concrete structs here and instead using interfaces. In that world, you'd have a creature interface, and both human and alien would satisfy the creature interface. You wouldn't necessarily know which one you were dealing with downstream, but you'd be able to reliably call creature methods and the appropriate human or alien implementation would be invoked.

Why can't enum variants be inferred in match arms?

This example fails to compile:
extern crate nix;
use std::os::unix::io::RawFd;
fn func(fd: RawFd, buf: &mut [u8]) -> Result<(), nix::Error> {
let (size, nix_addr) = nix::sys::socket::recvfrom(
fd, buf
)?;
let addr = match nix_addr {
//nix::sys::socket::SockAddr::Inet(addr) => addr.to_std(),
Inet(addr) => addr.to_std(),
_ => panic!(),
};
Ok(())
}
fn main() {}
The error, in this version is:
error[E0531]: unresolved tuple struct/variant `Inet`
--> match_arms.rs:14:3
|
14 | Inet(addr) => addr.to_std(),
| ^^^^
Swapping the Inet line for the commented out one successfully compiles.
The compiler seems to be requiring me to specify the enum type itself, I suppose so that it knows that the variant I'm specifying in the match arm is legit. But why? Can't the enum be inferred? Doesn't the compiler have enough information here to realize that nix_addr is a nix::…::SocketAddr, and thus, that Inet is a valid variant (and one with data)?
Why do I have to type the whole thing out, or drag the name into the current scope with a use?
I also tried _::Inet, which also failed.
But why? Can't the enum be inferred?
According to RFC 390 which introduced enum namespacing, this inference is considered a hack and not having it is better designwise. From the RFC's alternatives section:
We can implement enum namespacing after 1.0 by adding a "fallback" case to resolve, where variants can be referenced from their "flat" definition location if no other definition would conflict in that namespace. In the grand scheme of hacks to preserve backwards compatibility, this is not that bad, but still decidedly worse than not having to worry about fallback at all.
The official reason why inference is not considered after RFC 390, is no one really cared enough to propose the change:
#sfackler:
#netvl Java's an interesting case in that you can only refer to variants in the "bare" form (FOO, not MyEnum.FOO) in switch statements. The situation in Rust is a bit more complex as match allows more powerful pattern matching than traditional switch statements. The closest analogue would probably be to implicitly treat all the relevant stuff as imported in a pattern. That seems like something sufficiently orthogonal to this proposal that it'd probably deserve its own RFC.
(and no one has written an RFC for this since then.)
The chance such an RFC will pass is slim, though. After all, you just need to add one line use nix::sys::socket::SockAddr::* somewhere to make it work. Adding a feature to the language requires so many consideration about the proper specification and corner cases (e.g. what happens when you use nix::sys::socket::SockAddr::Unix as Inet) that it may not worth the time.

What does :: mean in Rust?

What does the :: syntax in Rust, as seen here, mean:
fn chunk(n: uint, idx: uint) -> uint {
let sh = uint::BITS - (SHIFT * (idx + 1));
(n >> sh) & MASK
}
In languages like Haskell it means a type hint, but here the compiler already has an annotation of that values type, so it seems it's likely type casting.
Please review Appendix B: Operators and Symbols of The Rust Programming Language.
In this case, the double colon (::) is the path separator. Paths are comprised of crates, modules, and items.
The full path for your example item, updated for 1.0 is:
std::usize::BITS
Here, std is the crate, usize is a module, and BITS is the specific item — in this case a constant.
If you scroll up in your file, you'll see use core::usize. use adds the path to the set of items to look in. That's how you can get away with just saying usize::BITS. The core crate is an implementation detail of the façade that is the std crate, so you can just substitute std for core in normal code.
:: can also be used as a way to specify generic types when they cannot otherwise be inferred; this is called the turbofish.
See also:
What is the syntax: `instance.method::<SomeThing>()`?
Oops. I wasn't reading very clearly. In this case, it's just the normal way of referring to anything under a module. uint::BITS is a constant, it seems.

Rationale for Go's method syntax

Ok, I have to admit, I don't really use Go very much at all, but I did just observe something that strikes me as odd for a language that strives for minimality and all that good stuff as Go does. I would be surprised if there isn't a legitimate rationale behind it, so that's what I'm looking for.
So when you have a method, you define it like this:
func (s *SomeStruct) Foo(x int) { }
but why have an extra parameter list just for the "receiver", as I think it's called? Would it not be a simpler and more elegant design to just do
func Foo(s *SomeStruct, x int) { }
and then have s.Foo(5) just be translated to a call to a function Foo(s, 5)?
Methods are fundamentally special and different from regular functions.
Methods must live in the same package as the receiver type.
Methods are used to satisfy interfaces.
Receiver parameters are the only parameters which may be overloaded.
When anonymous struct fields have methods, those methods are "inherited".
With your proposal, the line between a function and a method becomes very blurry and it's difficult to figure out how to resolve the above issues.
That said, I think it would be really interesting to design a language with multimethods and interfaces. However, that language would not be Go.
Your question correctly points out that any method is a function. However, the Go language needs to be able to explicitly distinguish between methods and functions. The reason for this is that methods have features that functions do not have. The choice of whether Foo is a function or a method needs to be made by the programmer.
Go's minimalism means that the language defines only a small set of keywords. The Go authors could have chosen a new keyword, such as method, to distinguish methods from functions:
method Foo(receiver *T, arg1 int) {} // 'Foo' is a method, not a function
Looking around the Go programming language, we can see that the philosophy is to reuse already existing keywords rather than to have a separate keyword for each occasion. The for keyword is a good example of this approach:
for {} // Infinite loop
for a>0 {a--} // A while-do loop
for i := range channel {} // Receive values from a channel
for i:=0; i<N; i++ {} // C-style for loop
The basic idea is that for the parser (and Go programmers) to distinguish the various types of for loops from each other, there is no need to introduce a new keyword if the options can be distinguished by the syntax of what comes after the for keyword: ; := range identifier ..., their sequential order, and their presence/absence.
The func keyword follows the same pattern. It can be used in multiple contexts:
function definitions: func f() {}
function types: type F func(int) int
method definitions: func (t T) SomeMethod() {}
closures: { ...; go func(){c<-1}(); ...}
From minimalism viewpoint, a single func keyword is both simpler and more elegant than having multiple keywords.
The extra parameter list just for the receiver
func (t *T) Foo(x int) {}
enables the parser to distinguish methods and functions:
func IDENTIFIER ... This is going to be a function
func ( ... This is going to be a method
Thus, the parser (as well as Go programmers) can make the distinction based on whether the func keyword is followed by an identifier or by (.
The proposed replacement is not semantically identical to the current state, i.e. it's not a syntactic change only. It will [attempt to] automagically create methods of any [local package] type that happens to be the first parameter of a function. Considering how fundamental method sets are wrt to Go's concept of automatic interface satisfaction rules, this will quite probably lead to a big big mess.
In short, I think such change to the Go language improves nothing while damaging a lot of it's nice features related to methods and interfaces.
Probably because go isn't Python.
Also, because a function declared this way is not a method.
You cannot declare a method outside the object package. So I guess the main rationale in the syntax difference between methods and functions is to be able to differentiate them.

Resources