Views with repeating indices lead to different results than Base.Array

[clintonTE]

Describe the bug
Views don't play well with assignments, leading to silent errors.

To Reproduce
The Minimal Working Example (MWE) for this bug:

using Revise, CuArrays, Flux
function cusubmwe(N)
  v = ones(Float32, N)
  vgpu = v|> gpu

  indices = (i->(i-1)%N + 1).(1:N^2)
  subv = view(v, indices)
  subvgpu = view(vgpu, indices)

  subv .+= 1
  subvgpu .+= 1

  @info "indices: $indices"
  @info "cpu: $(v)"
  @info "gpu: $(vgpu)"

end

Output:

[ Info: indices: [1, 2, 3, 1, 2, 3, 1, 2, 3]
[ Info: cpu: Float32[4.0, 4.0, 4.0]
[ Info: gpu: Float32[2.0, 2.0, 2.0]

Expected behavior
The cpu version provides a baseline.

Build log

   Building WebIO ──→ `C:\Users\Clinton\.julia\packages\WebIO\2mZPb\deps\build.log`
   Building NNlib ──→ `C:\Users\Clinton\.julia\packages\NNlib\FAI3o\deps\build.log`
   Building Libtask → `C:\Users\Clinton\.julia\packages\Libtask\GQPaW\deps\build.log`
   Building CMake ──→ `C:\Users\Clinton\.julia\packages\CMake\ULbyn\deps\build.log`
   Building NLopt ──→ `C:\Users\Clinton\.julia\packages\NLopt\eqN9a\deps\build.log`

Environment details (please complete this section)
Details on Julia:

Commit b8e9a9ecc6 (2020-03-21 16:36 UTC)
Platform Info:
  OS: Windows (x86_64-w64-mingw32)
  CPU: Intel(R) Core(TM) i7-9750H CPU @ 2.60GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-8.0.1 (ORCJIT, skylake)
Environment:
  JULIA_EDITOR = "C:\Users\Clinton\AppData\Local\atom\app-1.45.0\atom.exe"  -a
  JULIA_NUM_THREADS = 6

Julia packages:
CuArrays v2.0.1 #master (https://github.com/JuliaGPU/CuArrays.jl.git)
GPUArrays v3.1.0 #master (https://github.com/JuliaGPU/GPUArrays.jl.git)
CUDAnative v3.0.4
Flux v0.10.4 #master (https://github.com/FluxML/Flux.jl.git)

CUDA: toolkit and driver version
10.2 / 441.66

Additional context
This came up in the context of FluxML/Zygote.jl#600, but I am basing the mwe on an example from Michael Abbott from Slack (https://julialang.slack.com/archives/C6G240ENA/p1587070650391500).

I think this is the reason I was having difficulties with Zygote AD on views of CuArrays as mentioned in that issue, though I don't know for sure.

[CarloLucibello]

Any hint on how to fix this?

[maleadt]

Views with repeating indices cannot be realistically supported without serializing the computation and losing all benefit of the GPU.