[swift-evolution] [swift-evolution-announce] [Re-Review] SE-0104: Protocol-oriented integers

[Dave Abrahams]

Sent from my moss-covered three-handled family gradunza

> On Feb 21, 2017, at 10:08 AM, John McCall <rjmccall at apple.com> wrote:
> 
> 
>> On Feb 21, 2017, at 2:15 PM, Dave Abrahams via swift-evolution <swift-evolution at swift.org> wrote:
>> 
>> 
>> 
>> Sent from my moss-covered three-handled family gradunza
>> 
>>> On Feb 21, 2017, at 9:04 AM, Jordan Rose <jordan_rose at apple.com> wrote:
>>> 
>>> [Proposal: https://github.com/apple/swift-evolution/blob/master/proposals/0104-improved-integers.md]
>>> 
>>> Hi, Max (and Dave). I did have some questions about this revision:
>>> 
>>>> Arithmetic and SignedArithmetic protocols have been renamed to Number and SignedNumber.
>>> 
>>> What happens to NSNumber here? It feels like the same problem as Character and (NS)CharacterSet.
>>> 
>>> 
>>>> Endian-converting initializers and properties were added to the FixedWidthInteger protocol.
>>> 
>>> This is the thing I have the biggest problem with. Endian conversions aren't numeric operations, and you can't meaningfully mix numbers of different endianness. That implies to me that numbers with different endianness should have different types. I think there's a design to explore with LittleEndian<Int> and BigEndian<Int>, and explicitly using those types whenever you need to convert.
>> 
>> I disagree. Nobody actually wants to compute with numbers in the wrong endianness for the machine. This is just used for corrections at the ends of wire protocols, where static type has no meaning.
> 
> I think Jordan's suggestion is not that LittleEndian<Int> or BigEndian<Int> would be artihmetic types, but that they would be different types, primarily opaque, that can be explicitly converted to/from Int.  When you read something off the wire, you ask for the bytes as one of those two types (as appropriate) and then convert to the underlying type.  Ideally, Int doesn't even conform to the "this type can be read off the wire" protocol, eliminating the common mistake of serializing something using native endianness.

Still, you have to implement those somehow. How do you do that without this functionality in the Integer API?  Turtles have to stop somewhere. We could de-emphasize these APIs by making them static, but "x.yyyy" already is a place of reduced emphasis for integers. 

> John.
> 
>> 
>>> Here's a sketch of such a thing:
>>> 
>>> struct LittleEndian<Value: FixedWidthInteger> {
>>>   private var storage: Value
>>> 
>>>   public var value: Value {
>>> #if little_endian
>>>     return storage
>>> #else
>>>     return swapBytes(storage)
>>> #endif
>>>   }
>>> 
>>>   public var bitPattern: Value {
>>>     return storage
>>>   }
>>> 
>>>   public var asBigEndian: BigEndian<Value> {
>>>     return BigEndian(value: self.value)
>>>   }
>>> 
>>>   public init(value: Value) {
>>> #if little_endian
>>>     storage = value
>>> #else
>>>     storage = swapBytes(value)
>>> #endif
>>>   }
>>> 
>>>   public init(bitPattern: Value) {
>>>     storage = bitPattern
>>>   }
>>> }
>>> 
>>> I'm not saying this is the right solution, just that I suspect adding Self-producing properties that change endianness is the wrong one.
>>> 
>>>>   /// The number of bits equal to 1 in this value's binary representation.
>>>>   ///
>>>>   /// For example, in a fixed-width integer type with a `bitWidth` value of 8,
>>>>   /// the number 31 has five bits equal to 1.
>>>>   ///
>>>>   ///     let x: Int8 = 0b0001_1111
>>>>   ///     // x == 31
>>>>   ///     // x.popcount == 5
>>>>   var popcount: Int { get
>>>>  }
>>> 
>>> Is this property actually useful enough to put into a protocol? I know it's defaulted, but it's already an esoteric operation; it seems unlikely that one would need it in a generic context. (It's also definable for arbitrary UnsignedIntegers as well as arbitrary FixedWidthIntegers.)
>> 
>> The whole point is that you want to dispatch down to an LLVM instruction for this and not rely on the optimizer to collapse your loop into one. 
>> 
>>> 
>>> (I'm also still not happy with the non-Swifty name, but I see "populationCount" or "numberOfOneBits" would probably be worse.)
>>> 
>>> 
>>> Thanks in advance,
>>> Jordan
>> _______________________________________________
>> swift-evolution mailing list
>> swift-evolution at swift.org
>> https://lists.swift.org/mailman/listinfo/swift-evolution
> 
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.swift.org/pipermail/swift-evolution/attachments/20170221/dcdfcbc5/attachment.html>