`@Const` on a reshaped view fails to compile on CUDA

[maleadt]

A kernel whose @Const argument is a ReshapedArray (e.g. vec(view(A, 1:4, 1:4))) fails to compile on CUDA; without @Const the same kernel works.

using CUDA, KernelAbstractions

@kernel function copy_const!(y, @Const(x))
    i = @index(Global)
    @inbounds y[i] = x[i]
end
@kernel function copy_plain!(y, x)
    i = @index(Global)
    @inbounds y[i] = x[i]
end

A = CuArray(Float32.(reshape(1:25, 5, 5)))
x = vec(view(A, 1:4, 1:4))                 # a ReshapedArray of a SubArray
y = CuArray{Float32}(undef, 16)
copy_plain!(CUDABackend())(y, x; ndrange=16)   # works
copy_const!(CUDABackend())(y, x; ndrange=16)   # InvalidIRError

InvalidIRError: compiling MethodInstance for gpu_copy_const!(..., ::CuDeviceVector{Float32, 1}, ::Base.ReshapedArray{Float32, 1, SubArray{Float32, 2, CuDeviceMatrix{Float32, 1}, Tuple{UnitRange{Int64}, UnitRange{Int64}}, false}, Tuple{Base.MultiplicativeInverses.SignedMultiplicativeInverse{Int64}}}) resulted in invalid LLVM IR
Reason: unsupported call to a lazy-initialized function (call to ijl_rethrow)
Stacktrace:
 [1] rethrow                       @ ./error.jl:71
 ...
 [5] SignedMultiplicativeInverse   @ ./multinverses.jl:54
 ...
 [10] reshape                      @ ./reshapedarray.jl:128
 [11] adapt_structure              @ Adapt/src/wrappers.jl:18
 [12] adapt                        @ Adapt/src/Adapt.jl:40
 [13] adapt_structure              @ Adapt/src/wrappers.jl:10
 [14] adapt                        @ Adapt/src/Adapt.jl:40
 [15] constify                     @ KernelAbstractions/src/KernelAbstractions.jl:464

KernelAbstractions 0.9.42 (and main's constify is the same), CUDA.jl 6.4.0, Julia 1.13.0, RTX 5080.

@Const(x) expands to x = constify(x) inside the kernel, and constify(x) = adapt(ConstAdaptor(), x). For a ReshapedArray, Adapt rebuilds the wrapper with reshape(adapt(to, parent(x)), size(x)), which recomputes the SignedMultiplicativeInverses; their constructor has an error path that builds a string, which cannot compile for the GPU. So any @Const argument that is a reshape (of a view, or of anything that is not a plain device array) fails. Possible fixes: rebuild the ReshapedArray without recomputing its inverses (e.g. Base.ReshapedArray(adapt(to, parent(x)), size(x), x.mi)) in an adapt_structure method for ConstAdaptor, or apply the const marking only to the innermost device array.

AcceleratedKernels hits this in reduce/count of such views, which GPUArrays now routes to it (for example GPUArrays' norm(view(x, ...), 0) tests on CuArray). Its rework (JuliaGPU/AcceleratedKernels.jl#133) works around it by applying @Const only to plain dense device arrays, and will drop that once this is fixed.