[swift-evolution] [Proposal] Introduces endianness specific type

Robert Bennett rltbennett at icloud.com
Sun Jul 9 17:57:57 CDT 2017


Just a question: how would/does allowing the reordering of fields affect the correctness and performance of the (de)serialization API added in Swift 4?

> On Jul 9, 2017, at 6:21 PM, Jens Persson via swift-evolution <swift-evolution at swift.org> wrote:
> 
> Thanks for that clarification John McCall.
> My code is using a lot of generic structs (in which memory layout is important) though, an example would be:
> struct Vector4<E> : Vector {
>     typealias Index = VectorIndex4
>     typealias Element = E
>     var e0, e1, e2, e3: Element
>> }
> And I guess I'm out of luck trying to represent those as C structs?
> So AFAICS it looks like it is currently impossible to write generic low level code in Swift, unless I just keep doing what I've been doing (It does currently work after all) knowing that it will probably break in some future versions of Swift. But in that possible future version of Swift, I could probably find a way to make it work again (using some possible explicit tools for layout control present in that version of Swift).
> Correct?
> /Jens
> 
> 
>> On Sun, Jul 9, 2017 at 11:41 PM, John McCall <rjmccall at apple.com> wrote:
>> 
>>> On Jul 9, 2017, at 4:49 PM, Jens Persson via swift-evolution <swift-evolution at swift.org> wrote:
>>> 
>>> Sorry for making so much off topic noise in this thread, but I made a mistake regarding the Metal tutorial:
>>> Looking more carefully I see now that they rebuild a vertedData: [Float] from their vertices: [Vertex] using the floatBuffer() method of the Vertex struct, which returns an Array with the stored properties of Vertex in correct order.
>>> 
>>> While searching the internet about this I saw Joe Groff mentioning on Twitter that:
>>> "We plan to sort fields in padding order to minimize size, and may also automatically pack bools and enums in bitfields."
>>> 
>>> So AFAICS my current image processing code is making the possibly invalid assumption that eg
>>> struct S {
>>>     var a, b, c, d: Float
>>> }
>>> will have a memory layout of 4*4=16 bytes (stride and size == 16) and an alignment of 4, and most importantly that a, b, c, d will be in that order.
>> 
>> This is currently true, but may not always be.  We want to reserve the right to re-order the fields even if it doesn't improve packing — for example, if two fields are frequently accessed together, field reordering could yield substantial locality benefits.  We've also discussed reordering fields to put them in a canonical order for resilience, since it's a little counter-intuitive that reordering the fields of a struct should be ABI-breaking.  (There are arguments against doing that as well, of course — for example, the programmer may have chosen the current order for their own locality optimizations.)
>> 
>>> It looks like I should be defining my structs (the ones for which memory layout is important) in C and import them.
>> 
>> This should always work, yes.
>> 
>>> Although I would be surprised if a Swift-struct containing only same-sized fields (all of the same primitive type) would be reordered, and such changes to the language would probably include some per-struct way to express some sort of layout control (since being able to define structs to be used for low level data manipulation is important in a systems language).
>> 
>> Exactly.  In the long term, Swift will have some explicit tools for layout control.
>> 
>> John.
>> 
>>> 
>>> /Jens
>>> 
>>> 
>>>> On Sun, Jul 9, 2017 at 7:01 PM, Jens Persson via swift-evolution <swift-evolution at swift.org> wrote:
>>>> I should perhaps add that in my image processing code, I use code like this:
>>>> 
>>>> func withVImageBuffer<Data, R>(for table: Table<Data>, body: (vImage_Buffer) -> R) -> R
>>>>     where
>>>>     Data.Coordinate.Index == VectorIndex2
>>>> {
>>>>     let vib = vImage_Buffer(
>>>>         data: table.baseAddress,
>>>>         height: vImagePixelCount(table.size.e1),
>>>>         width: vImagePixelCount(table.size.e0),
>>>>         rowBytes: table.stride.e1
>>>>     )
>>>>     return withExtendedLifetime(table) { body(vib) }
>>>> }
>>>> 
>>>> Here, Table<Data> is the raster image. Data.Coordinate == VectorIndex2 makes it a 2D table, and a Table's Data also has a type parameter Data.Value which can be eg one of the "pixel"-struct I showed before.
>>>> This works without any problems (I've tested and used the some variant of this type of code for years) but it would surely break if the memory layout of simple structs changed.
>>>> 
>>>> I can't see how this usage is much different from the one in the Metal tutorial. It too uses pointers to point into a data created using the (Swift) struct "Vertex", and the GPU hardware has its expectations on the memory layout of that data, so the code would break if the memory layout of the Vertex struct changed.
>>>> 
>>>> /Jens
>>>> 
>>>> 
>>>>> On Sun, Jul 9, 2017 at 6:35 PM, Jens Persson <jens at bitcycle.com> wrote:
>>>> 
>>>>> I don't think I'm misunderstanding you, but I might be, so I'll add more detail:
>>>>> 
>>>>> If you look at the Metal article, you'll see that the (Swift) struct "Vertex" is used to specify the data that is sent to Metal for creating a buffer (using MTLDevice.makeBuffer). The result that the GPU will produce surely depends on the fields of the Vertex struct (x, y, z, r, g, b, a) being in the specified order (ie swapping the red channel with the x coordinate would produce an unexpected result).
>>>>> 
>>>>> And regarding the second example, pixel structs used for manipulating raster image data. Manipulating raster image data presumably includes stuff like displaying to screen, loading and saving raster images.
>>>>> I currently use this way of doing this right now without any problems, but if the order of the fields (eg a, r, g, b) should change in the future, then my code would break (the colors of the images would at least not come out as expected).
>>>>> 
>>>>> /Jens
>>>>> 
>>>>> 
>>>>> 
>>>>> 
>>>>> 
>>>>> 
>>>>>> On Sun, Jul 9, 2017 at 5:53 PM, Chris Lattner <clattner at nondot.org> wrote:
>>>>>> 
>>>>>>> On Jul 9, 2017, at 12:23 AM, Jens Persson <jens at bitcycle.com> wrote:
>>>>>>> 
>>>>>>> 
>>>>>>>> On Sat, Jul 8, 2017 at 6:28 PM, Chris Lattner via swift-evolution <swift-evolution at swift.org> wrote:
>>>>>>>> Hi Susan,
>>>>>>>> 
>>>>>>>> Swift does not currently specify a layout for Swift structs.  You shouldn’t be using them for memory mapped i/o or writing to a file, because their layout can change.  When ABI stability for fragile structs lands, you will be able to count on it, but until then something like this is probably a bad idea.
>>>>>>>> 
>>>>>>>> -Chris
>>>>>>> 
>>>>>>> Does this imply that you should never use Swift structs to eg interact with Metal?
>>>>>> 
>>>>>> No.
>>>>>> 
>>>>>>> This seems to be a very common practice. Here is a typical example (from a Metal tutorial at raywenderlich.com):
>>>>>>> 
>>>>>>> struct Vertex {
>>>>>>>   var x,y,z: Float     // position data
>>>>>>>   var r,g,b,a: Float   // color data
>>>>>>> 
>>>>>>>   func floatBuffer() -> [Float] {
>>>>>>>     return [x,y,z,r,g,b,a]
>>>>>>>   }
>>>>>>> }
>>>>>> 
>>>>>> This doesn’t appear to expose the layout of the struct.
>>>>>>> Also, does it imply that we cannot use structs (of only primitive types) like:
>>>>>>> 
>>>>>>> struct RgbaFloatsLinearGamma {
>>>>>>>     var r, g, b, a: Float
>>>>>>>>>>>>>> }
>>>>>>> struct BgraBytesSrgbGamma {
>>>>>>>     var b, g, r, a: UInt8
>>>>>>> }
>>>>>>> 
>>>>>>> for manipulating raster image data?
>>>>>> 
>>>>>> I don’t see why that would be a problem.
>>>>>> 
>>>>>>> I vaguely remember a swift evo discussion where it was concluded that such usage was considered OK provided the stored properties of the structs was only primitive types, but I can't find it now.
>>>>>>> 
>>>>>>> Perhaps it could be considered OK at least when the intended platforms are known to be only iOS devices?
>>>>>> 
>>>>>> I think you’re misunderstanding what I’m saying.  It isn’t correct to take (e.g.) an unsafepointer to the beginning of a struct, and serialize that out to disk, and expect that the fields are emitted in some order with some specific padding between them.  None of the uses above try to do this.
>>>>>> 
>>>>>> -Chris
>>>>>> 
>>>>>> 
>>>>> 
>>>> 
>>>> 
>>>> _______________________________________________
>>>> swift-evolution mailing list
>>>> swift-evolution at swift.org
>>>> https://lists.swift.org/mailman/listinfo/swift-evolution
>>>> 
>>> 
>>> _______________________________________________
>>> swift-evolution mailing list
>>> swift-evolution at swift.org
>>> https://lists.swift.org/mailman/listinfo/swift-evolution
>> 
> 
> _______________________________________________
> 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/20170709/c09cb691/attachment.html>


More information about the swift-evolution mailing list