[swift-evolution] [pitch] remove global sqrt<T: FloatingPoint>

G B g.c.b.at.work at gmail.com
Mon Jul 11 17:49:22 CDT 2016


I understand that it’s only one function and that I’m probably being ever-so-slightly unreasonable but I’ve spent the last month going crazy that a language as expressive as Swift, with all of its support for protocols and generics and under-the-hood optimization, makes me do stuff like this to abstract the most typical of algorithms across functionally identical types:

protocol ScalarMathType  {
    func sqrt() -> Self
    func sin() -> Self
    func cos() -> Self
    func asin() -> Self
    func acos() -> Self
    func atan2(x:Self) -> Self
}


func sqrt<T:ScalarMathType> (x:T) -> T {return x.sqrt()}
func sin<T:ScalarMathType> (x:T) -> T {return x.sin()}
func cos<T:ScalarMathType> (x:T) -> T {return x.cos()}
func asin<T:ScalarMathType>(x:T) -> T {return x.asin()}
func acos<T:ScalarMathType>(x:T) -> T {return x.acos()}
func atan2<T:ScalarMathType>(y:T,x:T) -> T {return y.atan2(x)}



extension Float  : ScalarMathType {
    func sqrt() -> Float {return Foundation.sqrt(self)}
    func sin() -> Float {return Foundation.sin(self)}
    func cos() -> Float {return Foundation.cos(self)}
    func asin() -> Float {return Foundation.asin(self)}
    func acos() -> Float {return Foundation.acos(self)}
    func atan2(x:Float) -> Float {return Foundation.atan2(self,x)}
}

extension Double : ScalarMathType {
    func sqrt() -> Double {return Foundation.sqrt(self)}
    func sin() -> Double {return Foundation.sin(self)}
    func cos() -> Double {return Foundation.cos(self)}
    func asin() -> Double {return Foundation.asin(self)}
    func acos() -> Double {return Foundation.acos(self)}
    func atan2(x:Double) -> Double {return Foundation.atan2(self,x)}
}


That just seems so horribly hackish.

Even if the `Math` module only included this, it would be a win.  Otherwise, I’d almost prefer that we stuff these into the global namespace to keep `sqrt` company.  Right now every developer who wants to change between polar and cartesian coordinates, or calculate a radius from a circumference in a generic way needs to replicate this code…  

Hopefully we all do it the same way and in the same way that Swift4 implements it or we’re all going to have to deal with the conflicts later.



> On Jul 11, 2016, at 15:28 , Stephen Canon <scanon at apple.com> wrote:
> 
> The problem with having the Math module concurrent with this change is that such a thing is necessarily post-swift 3, whereas this change should happen in Swift 3 if it’s going to happen at all.
> 
> Given that this only effects one function, and you already need your own wrappers for every other math operation, it seems like the right tradeoff to do it now.
> 
> – Steve
> 
>> On Jul 11, 2016, at 6:20 PM, G B <g.c.b.at.work at gmail.com> wrote:
>> 
>> For certain functions (sqrt, sin, cos, log, etc) we’ve grown up in math class calling them as a function on an argument rather than viewing them as a property (or method) of the number.
>> 
>> Just like we prefer to use `a - b` rather than `a.sub(b)`, we (or at least I) prefer `sqrt(a)` over `a.squareRoot()`.
>> 
>> It seems this is dooming everyone to either reverse their internal math grammar or do what I’m doing now and write their own generic function for no other purpose than to wrap a method call to look like a function call.
>> 
>> I’d be less resistant if your proposed `Math` module was concurrent with the change.
>> 
>>> On Jul 11, 2016, at 15:13 , Stephen Canon <scanon at apple.com> wrote:
>>> 
>>> It wouldn’t delay code intended to operate generically at all.  That code can use .squareRoot( ).
>>> 
>>>> On Jul 11, 2016, at 6:12 PM, G B <g.c.b.at.work at gmail.com> wrote:
>>>> 
>>>> While I don’t have a strong opinion about what functions are in the global namespace and which are in a `Math` module, I’m not excited about the idea of delaying the availability of generic implementations of floating point functions.
>>>> 
>>>> How would this affect code intended to operate generically over Float and Double?  I’ve made the mistake of trying to do this with some of my code and it’s remarkably painful for what I’d hoped would be a simple abstraction.
>>>> 
>>>> Right now (pre SE-0067), it takes a surprising amount of tinkering to get code to work generically across those two types.  Provisions need to be added to provide `sqrt`, `sin`, `cos`, etc.  While it all compiles down to the same instructions, I don’t feel it is natural to call `squareRoot()` as a method.
>>>> 
>>>> I don’t necessarily care if these functions are in the global namespace, or if they’re imported from a `Math` module.  I’m also not convinced that they should be part of the core FloatingPoint protocol.  `sqrt` probably should be, but the trig functions would naturally fit together in a protocol that itself conforms to FloatingPoint.
>>>> 
>>>>> On Jul 11, 2016, at 14:28 , Stephen Canon via swift-evolution <swift-evolution at swift.org> wrote:
>>>>> 
>>>>> Post SE-0067 FloatingPoint provides the usual global operators, as well as a single global function:
>>>>> 
>>>>> 	func sqrt<T: FloatingPoint>(_: T) -> T
>>>>> 
>>>>> It seems out of place and lonely, and it would be nice if we can keep the default members of the global namespace to a minimum.
>>>>> 
>>>>> I’d like to suggest removing this global from FloatingPoint while keeping the existing global functions for concrete types in the Darwin.C module.  The square root operation would still be available for all FloatingPoint types as `.squareRoot()`.
>>>>> 
>>>>> I would also plan to provide this and other math.h-ish globals in a future (post swift 3) Math module.
>>>>> 
>>>>> – Steve
> 



More information about the swift-evolution mailing list