the GenericMemory type parameters

[nsajko]

#53854 (comment) @oscardssmith

We know that regular Memory has the interface we want, but I think we wanted a release or two and a few examples using the addrspace and kind before we actually decide that the GenericMemory doesn't have any design defects that would require breaking changes if it was part of the interface.

The order of the parameters: could the element type be moved to the front?

GenericMemory has three type parameters, of which the element type is set in stone, as it's required by the AbstractVector supertype; the remaining two parameters seem subject to change. I think it would make sense to make the element type the first type parameter, and perhaps declare in the docs that the rest of the type parameters are experimental/subject to change. This would make it more convenient to use GenericMemory as a type parameter in user types while avoiding the non-public type parameters of GenericMemory.

For example, the FixedSizeArrays package (not registered yet), is currently backed by Memory storage, but it'd ideally support GenericMemory: JuliaArrays/FixedSizeArrays.jl#33. Currently the struct definition looks like this (ref):

struct Internal end

struct FixedSizeArray{T,N} <: DenseArray{T,N}
    mem::Memory{T}
    size::NTuple{N,Int}
    function FixedSizeArray{T,N}(::Internal, mem::Memory{T}, size::NTuple{N,Int}) where {T,N}
        new{T,N}(mem, size)
    end
end

GenericMemory support could look something like this:

struct FixedSizeArray{T,N,Mem<:GenericMemory{<:Any,T}} <: DenseArray{T,N}
    mem::Mem
    size::NTuple{N,Int}
    function FixedSizeArray{T,N,Mem}(
        ::Internal, mem::Mem, size::NTuple{N,Int},
    ) where {T,N,Mem<:GenericMemory{<:Any,T}}
        new{T,N,Mem}(mem, size)
    end
end

Moving the element type parameter of GenericMemory to the first place would enable avoiding the awkward <:Any above.

The kind parameter: could it be made a singleton type instead of a Symbol value?

Having something like this, unlike just Symbols, would be typo-proof:

module MemoryKind
  struct Atomic end
  struct NotAtomic end
end

const Memory = GenericMemory{T,MemoryKind.NotAtomic,CPU} where {T}

[oscardssmith]

The kind change seems reasonable to me. That was one of the later pieces added to the design , although one advantage of symbols is that the C code needs to create them so doing more complicated things is annoying.

The type parameter ordering was very deliberate. The reasoning is that most code that works with GenericMemory will be fairly generic on the eltype, but the atomic memory will usually require fairly different code because all the mutations need to supply a memory order.