reshape improvements

[c42f]

A question came up on slack, prompting me to realize that our reshape could be improved:

• With the vastly improved constant propagation in 1.0, we can have an efficient reshape without needing a Size wrapper. There's probably other operations for which this is true too.
• Base.reshape docs specify that the reshape result shares the underlying storage. Our reshape of MVector with Size doesn't seem to satisfy that, but it probably should.

[andyferris]

Interesting. The Base.reshape docs are probably referring to Array - it has special indirections and a backing buffer that none of the rest of us get to use, unfortunately, otherwise we would totally be using that for MArray. You kind of need that indirection so that b = reshape(a, ...) doesn't mutate a, shares the storage, and doesn't result in a b with more layers of indirection than a.

For MArray we could add new layer of indirection (e.g. a ReshapedArray wrapper). Would people want that more than just a new MVector? The performance cost for a large Array is O(N), but static arrays are meant to be tiny.

Regarding Size, if someone has the time to invest in seeing if it can be removed entirely, that would be cool.

[c42f]

Oh right. I was thinking reshape had been updated to return a wrapper for Base.Array, but I was getting it confused with reinterpret. Looks like reshape is still special and rewraps the underlying buffer, as you say.