Add insert/remove_axis_inplace for IxDyn arrays

[jturner314]

Without these methods, users must have ownership of dynamic-dimensional arrays to insert/remove axes, which is unnecessarily restrictive.

This is motivated by #528.

[bluss]

Nice!

Thinking more

[bluss]

I guess removing an axis can never make us lose the first element (at index [0, 0, ..]) right, so we don't need to recompute the array's pointer here?

[jturner314]

That's correct. We need to assert that the axis length is nonzero for a similar reason.

[bluss]

Remove is surprising, maybe just make it clear to the user that it's like an in place slice, we hide some elements and can't get them back. Weird for owned arrays, but we already have the same thing in slicing, so it makes sense.

[jturner314]

Remove is surprising, maybe just make it clear to the user that it's like an in place slice, we hide some elements and can't get them back.

I can certainly document this, but I agree it's still weird.

I've always thought that .remove_axis() on ArrayBase was a pretty weird method because the name doesn't clearly describe what it's doing. The fact that it panics when the axis has length 0 also makes me uncomfortable. .remove_axis(Axis(x)) is equivalent to .into_subview(Axis(x), 0) anyway, so I'm actually in favor of removing .remove_axis() entirely.

The question remains, though, what to do with .remove_axis_inplace() for IxDyn arrays. One idea is the following:

• Rename the existing .subview_inplace() method on ArrayBase to .narrow_axis() or something.
• Add a .subview_inplace() method for IxDyn arrays that actually removes the axis. (Don't add a .remove_axis_inplace() method.)

The problem with this is that changing .subview_inplace() like this is a breaking change in behavior that would cause problems that wouldn't be caught by the compiler, since both behaviors (keeping the axis or removing it) have the same types for IxDyn arrays.

A workaround for the difficult-to-catch breaking change issue is the following. I prefer this naming scheme anyway because it's more similar to the slice/slice_axis methods:

• Rename .subview() to .index_axis()
• Rename .subview_mut() to .index_axis_mut()
• Rename .subview_inplace() to .narrow_axis()/.select_one()/.select_single()/.select_index()/etc.
• Rename .into_subview() to .index_axis_move()
• Add .index_axis_inplace() to ArrayBase<S, IxDyn>

What do you think?

[bluss]

Sounds good! I guess index_axis makes more sense too (the major connection we have is to the axis iters that produce the same views).

[jturner314]

See #537, which does the renames mentioned above, except using .collapse_axis() instead of .narrow_axis() and additionally renaming .slice_inplace() to .slice_collapse().

[jturner314]

Closing in favor of #537.