Description of the problem

I've tried using the examples of 1-D slice iteration and (N-1)-dimensional iteration by replacing xt::xarray<int> with xt::xtensor<int, 3> and found unexpected outcomes:

• 1-D slice iteration (xt::axis_slice_begin1): compiles and run without errors, but all slice are empty (prints {}). Changing the axis does not change the outcome
• (N-1)-dimensional iteration: does not compile, fails with

  xtensor/xaxis_iterator.hpp:112:24: error: no matching constructor for
    initialization of 'shape_type' (aka 'array<unsigned long, 3UL>')
          shape_type shape(e_shape.size() - 1);
  

Desired outcome

Similar iteration as xt::xarray. Eventually with compile time check on the axis (given that shape is known).

Alternative considered

Casting back and forth to an xt::xarray. I am unsure of the performance impacts, if any.

System information

This is running xtensor 0.21.5 from conda-forge and compiling on MacOs with AppleClang 11.0.3